Loading...
1/*
2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
6 *
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21/*
22 * BITMAP UNPLUGGING:
23 *
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
26 * explanation.
27 *
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
32 * new additions.
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
39 * batch.
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
43 * miss any bits.
44 */
45
46#include <linux/blkdev.h>
47#include <linux/kthread.h>
48#include <linux/raid/pq.h>
49#include <linux/async_tx.h>
50#include <linux/module.h>
51#include <linux/async.h>
52#include <linux/seq_file.h>
53#include <linux/cpu.h>
54#include <linux/slab.h>
55#include <linux/ratelimit.h>
56#include "md.h"
57#include "raid5.h"
58#include "raid0.h"
59#include "bitmap.h"
60
61/*
62 * Stripe cache
63 */
64
65#define NR_STRIPES 256
66#define STRIPE_SIZE PAGE_SIZE
67#define STRIPE_SHIFT (PAGE_SHIFT - 9)
68#define STRIPE_SECTORS (STRIPE_SIZE>>9)
69#define IO_THRESHOLD 1
70#define BYPASS_THRESHOLD 1
71#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72#define HASH_MASK (NR_HASH - 1)
73
74static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
75{
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
78}
79
80/* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
85 * be valid.
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
88 */
89static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
90{
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
93 return bio->bi_next;
94 else
95 return NULL;
96}
97
98/*
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101 */
102static inline int raid5_bi_phys_segments(struct bio *bio)
103{
104 return bio->bi_phys_segments & 0xffff;
105}
106
107static inline int raid5_bi_hw_segments(struct bio *bio)
108{
109 return (bio->bi_phys_segments >> 16) & 0xffff;
110}
111
112static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113{
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
116}
117
118static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119{
120 unsigned short val = raid5_bi_hw_segments(bio);
121
122 --val;
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124 return val;
125}
126
127static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128{
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
130}
131
132/* Find first data disk in a raid6 stripe */
133static inline int raid6_d0(struct stripe_head *sh)
134{
135 if (sh->ddf_layout)
136 /* ddf always start from first device */
137 return 0;
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
140 return 0;
141 else
142 return sh->qd_idx + 1;
143}
144static inline int raid6_next_disk(int disk, int raid_disks)
145{
146 disk++;
147 return (disk < raid_disks) ? disk : 0;
148}
149
150/* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
154 */
155static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
157{
158 int slot = *count;
159
160 if (sh->ddf_layout)
161 (*count)++;
162 if (idx == sh->pd_idx)
163 return syndrome_disks;
164 if (idx == sh->qd_idx)
165 return syndrome_disks + 1;
166 if (!sh->ddf_layout)
167 (*count)++;
168 return slot;
169}
170
171static void return_io(struct bio *return_bi)
172{
173 struct bio *bi = return_bi;
174 while (bi) {
175
176 return_bi = bi->bi_next;
177 bi->bi_next = NULL;
178 bi->bi_size = 0;
179 bio_endio(bi, 0);
180 bi = return_bi;
181 }
182}
183
184static void print_raid5_conf (struct r5conf *conf);
185
186static int stripe_operations_active(struct stripe_head *sh)
187{
188 return sh->check_state || sh->reconstruct_state ||
189 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191}
192
193static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
194{
195 if (atomic_dec_and_test(&sh->count)) {
196 BUG_ON(!list_empty(&sh->lru));
197 BUG_ON(atomic_read(&conf->active_stripes)==0);
198 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 if (test_bit(STRIPE_DELAYED, &sh->state) &&
200 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
203 sh->bm_seq - conf->seq_write > 0)
204 list_add_tail(&sh->lru, &conf->bitmap_list);
205 else {
206 clear_bit(STRIPE_DELAYED, &sh->state);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
209 }
210 md_wakeup_thread(conf->mddev->thread);
211 } else {
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
214 if (atomic_dec_return(&conf->preread_active_stripes)
215 < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
217 atomic_dec(&conf->active_stripes);
218 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
219 list_add_tail(&sh->lru, &conf->inactive_list);
220 wake_up(&conf->wait_for_stripe);
221 if (conf->retry_read_aligned)
222 md_wakeup_thread(conf->mddev->thread);
223 }
224 }
225 }
226}
227
228static void release_stripe(struct stripe_head *sh)
229{
230 struct r5conf *conf = sh->raid_conf;
231 unsigned long flags;
232
233 spin_lock_irqsave(&conf->device_lock, flags);
234 __release_stripe(conf, sh);
235 spin_unlock_irqrestore(&conf->device_lock, flags);
236}
237
238static inline void remove_hash(struct stripe_head *sh)
239{
240 pr_debug("remove_hash(), stripe %llu\n",
241 (unsigned long long)sh->sector);
242
243 hlist_del_init(&sh->hash);
244}
245
246static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
247{
248 struct hlist_head *hp = stripe_hash(conf, sh->sector);
249
250 pr_debug("insert_hash(), stripe %llu\n",
251 (unsigned long long)sh->sector);
252
253 hlist_add_head(&sh->hash, hp);
254}
255
256
257/* find an idle stripe, make sure it is unhashed, and return it. */
258static struct stripe_head *get_free_stripe(struct r5conf *conf)
259{
260 struct stripe_head *sh = NULL;
261 struct list_head *first;
262
263 if (list_empty(&conf->inactive_list))
264 goto out;
265 first = conf->inactive_list.next;
266 sh = list_entry(first, struct stripe_head, lru);
267 list_del_init(first);
268 remove_hash(sh);
269 atomic_inc(&conf->active_stripes);
270out:
271 return sh;
272}
273
274static void shrink_buffers(struct stripe_head *sh)
275{
276 struct page *p;
277 int i;
278 int num = sh->raid_conf->pool_size;
279
280 for (i = 0; i < num ; i++) {
281 p = sh->dev[i].page;
282 if (!p)
283 continue;
284 sh->dev[i].page = NULL;
285 put_page(p);
286 }
287}
288
289static int grow_buffers(struct stripe_head *sh)
290{
291 int i;
292 int num = sh->raid_conf->pool_size;
293
294 for (i = 0; i < num; i++) {
295 struct page *page;
296
297 if (!(page = alloc_page(GFP_KERNEL))) {
298 return 1;
299 }
300 sh->dev[i].page = page;
301 }
302 return 0;
303}
304
305static void raid5_build_block(struct stripe_head *sh, int i, int previous);
306static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
307 struct stripe_head *sh);
308
309static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
310{
311 struct r5conf *conf = sh->raid_conf;
312 int i;
313
314 BUG_ON(atomic_read(&sh->count) != 0);
315 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
316 BUG_ON(stripe_operations_active(sh));
317
318 pr_debug("init_stripe called, stripe %llu\n",
319 (unsigned long long)sh->sector);
320
321 remove_hash(sh);
322
323 sh->generation = conf->generation - previous;
324 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 sh->sector = sector;
326 stripe_set_idx(sector, conf, previous, sh);
327 sh->state = 0;
328
329
330 for (i = sh->disks; i--; ) {
331 struct r5dev *dev = &sh->dev[i];
332
333 if (dev->toread || dev->read || dev->towrite || dev->written ||
334 test_bit(R5_LOCKED, &dev->flags)) {
335 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
336 (unsigned long long)sh->sector, i, dev->toread,
337 dev->read, dev->towrite, dev->written,
338 test_bit(R5_LOCKED, &dev->flags));
339 WARN_ON(1);
340 }
341 dev->flags = 0;
342 raid5_build_block(sh, i, previous);
343 }
344 insert_hash(conf, sh);
345}
346
347static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
348 short generation)
349{
350 struct stripe_head *sh;
351 struct hlist_node *hn;
352
353 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
354 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
355 if (sh->sector == sector && sh->generation == generation)
356 return sh;
357 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
358 return NULL;
359}
360
361/*
362 * Need to check if array has failed when deciding whether to:
363 * - start an array
364 * - remove non-faulty devices
365 * - add a spare
366 * - allow a reshape
367 * This determination is simple when no reshape is happening.
368 * However if there is a reshape, we need to carefully check
369 * both the before and after sections.
370 * This is because some failed devices may only affect one
371 * of the two sections, and some non-in_sync devices may
372 * be insync in the section most affected by failed devices.
373 */
374static int calc_degraded(struct r5conf *conf)
375{
376 int degraded, degraded2;
377 int i;
378
379 rcu_read_lock();
380 degraded = 0;
381 for (i = 0; i < conf->previous_raid_disks; i++) {
382 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
383 if (rdev && test_bit(Faulty, &rdev->flags))
384 rdev = rcu_dereference(conf->disks[i].replacement);
385 if (!rdev || test_bit(Faulty, &rdev->flags))
386 degraded++;
387 else if (test_bit(In_sync, &rdev->flags))
388 ;
389 else
390 /* not in-sync or faulty.
391 * If the reshape increases the number of devices,
392 * this is being recovered by the reshape, so
393 * this 'previous' section is not in_sync.
394 * If the number of devices is being reduced however,
395 * the device can only be part of the array if
396 * we are reverting a reshape, so this section will
397 * be in-sync.
398 */
399 if (conf->raid_disks >= conf->previous_raid_disks)
400 degraded++;
401 }
402 rcu_read_unlock();
403 if (conf->raid_disks == conf->previous_raid_disks)
404 return degraded;
405 rcu_read_lock();
406 degraded2 = 0;
407 for (i = 0; i < conf->raid_disks; i++) {
408 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
409 if (rdev && test_bit(Faulty, &rdev->flags))
410 rdev = rcu_dereference(conf->disks[i].replacement);
411 if (!rdev || test_bit(Faulty, &rdev->flags))
412 degraded2++;
413 else if (test_bit(In_sync, &rdev->flags))
414 ;
415 else
416 /* not in-sync or faulty.
417 * If reshape increases the number of devices, this
418 * section has already been recovered, else it
419 * almost certainly hasn't.
420 */
421 if (conf->raid_disks <= conf->previous_raid_disks)
422 degraded2++;
423 }
424 rcu_read_unlock();
425 if (degraded2 > degraded)
426 return degraded2;
427 return degraded;
428}
429
430static int has_failed(struct r5conf *conf)
431{
432 int degraded;
433
434 if (conf->mddev->reshape_position == MaxSector)
435 return conf->mddev->degraded > conf->max_degraded;
436
437 degraded = calc_degraded(conf);
438 if (degraded > conf->max_degraded)
439 return 1;
440 return 0;
441}
442
443static struct stripe_head *
444get_active_stripe(struct r5conf *conf, sector_t sector,
445 int previous, int noblock, int noquiesce)
446{
447 struct stripe_head *sh;
448
449 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
450
451 spin_lock_irq(&conf->device_lock);
452
453 do {
454 wait_event_lock_irq(conf->wait_for_stripe,
455 conf->quiesce == 0 || noquiesce,
456 conf->device_lock, /* nothing */);
457 sh = __find_stripe(conf, sector, conf->generation - previous);
458 if (!sh) {
459 if (!conf->inactive_blocked)
460 sh = get_free_stripe(conf);
461 if (noblock && sh == NULL)
462 break;
463 if (!sh) {
464 conf->inactive_blocked = 1;
465 wait_event_lock_irq(conf->wait_for_stripe,
466 !list_empty(&conf->inactive_list) &&
467 (atomic_read(&conf->active_stripes)
468 < (conf->max_nr_stripes *3/4)
469 || !conf->inactive_blocked),
470 conf->device_lock,
471 );
472 conf->inactive_blocked = 0;
473 } else
474 init_stripe(sh, sector, previous);
475 } else {
476 if (atomic_read(&sh->count)) {
477 BUG_ON(!list_empty(&sh->lru)
478 && !test_bit(STRIPE_EXPANDING, &sh->state));
479 } else {
480 if (!test_bit(STRIPE_HANDLE, &sh->state))
481 atomic_inc(&conf->active_stripes);
482 if (list_empty(&sh->lru) &&
483 !test_bit(STRIPE_EXPANDING, &sh->state))
484 BUG();
485 list_del_init(&sh->lru);
486 }
487 }
488 } while (sh == NULL);
489
490 if (sh)
491 atomic_inc(&sh->count);
492
493 spin_unlock_irq(&conf->device_lock);
494 return sh;
495}
496
497/* Determine if 'data_offset' or 'new_data_offset' should be used
498 * in this stripe_head.
499 */
500static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
501{
502 sector_t progress = conf->reshape_progress;
503 /* Need a memory barrier to make sure we see the value
504 * of conf->generation, or ->data_offset that was set before
505 * reshape_progress was updated.
506 */
507 smp_rmb();
508 if (progress == MaxSector)
509 return 0;
510 if (sh->generation == conf->generation - 1)
511 return 0;
512 /* We are in a reshape, and this is a new-generation stripe,
513 * so use new_data_offset.
514 */
515 return 1;
516}
517
518static void
519raid5_end_read_request(struct bio *bi, int error);
520static void
521raid5_end_write_request(struct bio *bi, int error);
522
523static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
524{
525 struct r5conf *conf = sh->raid_conf;
526 int i, disks = sh->disks;
527
528 might_sleep();
529
530 for (i = disks; i--; ) {
531 int rw;
532 int replace_only = 0;
533 struct bio *bi, *rbi;
534 struct md_rdev *rdev, *rrdev = NULL;
535 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
536 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
537 rw = WRITE_FUA;
538 else
539 rw = WRITE;
540 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
541 rw = READ;
542 else if (test_and_clear_bit(R5_WantReplace,
543 &sh->dev[i].flags)) {
544 rw = WRITE;
545 replace_only = 1;
546 } else
547 continue;
548 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
549 rw |= REQ_SYNC;
550
551 bi = &sh->dev[i].req;
552 rbi = &sh->dev[i].rreq; /* For writing to replacement */
553
554 bi->bi_rw = rw;
555 rbi->bi_rw = rw;
556 if (rw & WRITE) {
557 bi->bi_end_io = raid5_end_write_request;
558 rbi->bi_end_io = raid5_end_write_request;
559 } else
560 bi->bi_end_io = raid5_end_read_request;
561
562 rcu_read_lock();
563 rrdev = rcu_dereference(conf->disks[i].replacement);
564 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
565 rdev = rcu_dereference(conf->disks[i].rdev);
566 if (!rdev) {
567 rdev = rrdev;
568 rrdev = NULL;
569 }
570 if (rw & WRITE) {
571 if (replace_only)
572 rdev = NULL;
573 if (rdev == rrdev)
574 /* We raced and saw duplicates */
575 rrdev = NULL;
576 } else {
577 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
578 rdev = rrdev;
579 rrdev = NULL;
580 }
581
582 if (rdev && test_bit(Faulty, &rdev->flags))
583 rdev = NULL;
584 if (rdev)
585 atomic_inc(&rdev->nr_pending);
586 if (rrdev && test_bit(Faulty, &rrdev->flags))
587 rrdev = NULL;
588 if (rrdev)
589 atomic_inc(&rrdev->nr_pending);
590 rcu_read_unlock();
591
592 /* We have already checked bad blocks for reads. Now
593 * need to check for writes. We never accept write errors
594 * on the replacement, so we don't to check rrdev.
595 */
596 while ((rw & WRITE) && rdev &&
597 test_bit(WriteErrorSeen, &rdev->flags)) {
598 sector_t first_bad;
599 int bad_sectors;
600 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
601 &first_bad, &bad_sectors);
602 if (!bad)
603 break;
604
605 if (bad < 0) {
606 set_bit(BlockedBadBlocks, &rdev->flags);
607 if (!conf->mddev->external &&
608 conf->mddev->flags) {
609 /* It is very unlikely, but we might
610 * still need to write out the
611 * bad block log - better give it
612 * a chance*/
613 md_check_recovery(conf->mddev);
614 }
615 /*
616 * Because md_wait_for_blocked_rdev
617 * will dec nr_pending, we must
618 * increment it first.
619 */
620 atomic_inc(&rdev->nr_pending);
621 md_wait_for_blocked_rdev(rdev, conf->mddev);
622 } else {
623 /* Acknowledged bad block - skip the write */
624 rdev_dec_pending(rdev, conf->mddev);
625 rdev = NULL;
626 }
627 }
628
629 if (rdev) {
630 if (s->syncing || s->expanding || s->expanded
631 || s->replacing)
632 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
633
634 set_bit(STRIPE_IO_STARTED, &sh->state);
635
636 bi->bi_bdev = rdev->bdev;
637 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
638 __func__, (unsigned long long)sh->sector,
639 bi->bi_rw, i);
640 atomic_inc(&sh->count);
641 if (use_new_offset(conf, sh))
642 bi->bi_sector = (sh->sector
643 + rdev->new_data_offset);
644 else
645 bi->bi_sector = (sh->sector
646 + rdev->data_offset);
647 bi->bi_flags = 1 << BIO_UPTODATE;
648 bi->bi_idx = 0;
649 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
650 bi->bi_io_vec[0].bv_offset = 0;
651 bi->bi_size = STRIPE_SIZE;
652 bi->bi_next = NULL;
653 if (rrdev)
654 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
655 generic_make_request(bi);
656 }
657 if (rrdev) {
658 if (s->syncing || s->expanding || s->expanded
659 || s->replacing)
660 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
661
662 set_bit(STRIPE_IO_STARTED, &sh->state);
663
664 rbi->bi_bdev = rrdev->bdev;
665 pr_debug("%s: for %llu schedule op %ld on "
666 "replacement disc %d\n",
667 __func__, (unsigned long long)sh->sector,
668 rbi->bi_rw, i);
669 atomic_inc(&sh->count);
670 if (use_new_offset(conf, sh))
671 rbi->bi_sector = (sh->sector
672 + rrdev->new_data_offset);
673 else
674 rbi->bi_sector = (sh->sector
675 + rrdev->data_offset);
676 rbi->bi_flags = 1 << BIO_UPTODATE;
677 rbi->bi_idx = 0;
678 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
679 rbi->bi_io_vec[0].bv_offset = 0;
680 rbi->bi_size = STRIPE_SIZE;
681 rbi->bi_next = NULL;
682 generic_make_request(rbi);
683 }
684 if (!rdev && !rrdev) {
685 if (rw & WRITE)
686 set_bit(STRIPE_DEGRADED, &sh->state);
687 pr_debug("skip op %ld on disc %d for sector %llu\n",
688 bi->bi_rw, i, (unsigned long long)sh->sector);
689 clear_bit(R5_LOCKED, &sh->dev[i].flags);
690 set_bit(STRIPE_HANDLE, &sh->state);
691 }
692 }
693}
694
695static struct dma_async_tx_descriptor *
696async_copy_data(int frombio, struct bio *bio, struct page *page,
697 sector_t sector, struct dma_async_tx_descriptor *tx)
698{
699 struct bio_vec *bvl;
700 struct page *bio_page;
701 int i;
702 int page_offset;
703 struct async_submit_ctl submit;
704 enum async_tx_flags flags = 0;
705
706 if (bio->bi_sector >= sector)
707 page_offset = (signed)(bio->bi_sector - sector) * 512;
708 else
709 page_offset = (signed)(sector - bio->bi_sector) * -512;
710
711 if (frombio)
712 flags |= ASYNC_TX_FENCE;
713 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
714
715 bio_for_each_segment(bvl, bio, i) {
716 int len = bvl->bv_len;
717 int clen;
718 int b_offset = 0;
719
720 if (page_offset < 0) {
721 b_offset = -page_offset;
722 page_offset += b_offset;
723 len -= b_offset;
724 }
725
726 if (len > 0 && page_offset + len > STRIPE_SIZE)
727 clen = STRIPE_SIZE - page_offset;
728 else
729 clen = len;
730
731 if (clen > 0) {
732 b_offset += bvl->bv_offset;
733 bio_page = bvl->bv_page;
734 if (frombio)
735 tx = async_memcpy(page, bio_page, page_offset,
736 b_offset, clen, &submit);
737 else
738 tx = async_memcpy(bio_page, page, b_offset,
739 page_offset, clen, &submit);
740 }
741 /* chain the operations */
742 submit.depend_tx = tx;
743
744 if (clen < len) /* hit end of page */
745 break;
746 page_offset += len;
747 }
748
749 return tx;
750}
751
752static void ops_complete_biofill(void *stripe_head_ref)
753{
754 struct stripe_head *sh = stripe_head_ref;
755 struct bio *return_bi = NULL;
756 struct r5conf *conf = sh->raid_conf;
757 int i;
758
759 pr_debug("%s: stripe %llu\n", __func__,
760 (unsigned long long)sh->sector);
761
762 /* clear completed biofills */
763 spin_lock_irq(&conf->device_lock);
764 for (i = sh->disks; i--; ) {
765 struct r5dev *dev = &sh->dev[i];
766
767 /* acknowledge completion of a biofill operation */
768 /* and check if we need to reply to a read request,
769 * new R5_Wantfill requests are held off until
770 * !STRIPE_BIOFILL_RUN
771 */
772 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
773 struct bio *rbi, *rbi2;
774
775 BUG_ON(!dev->read);
776 rbi = dev->read;
777 dev->read = NULL;
778 while (rbi && rbi->bi_sector <
779 dev->sector + STRIPE_SECTORS) {
780 rbi2 = r5_next_bio(rbi, dev->sector);
781 if (!raid5_dec_bi_phys_segments(rbi)) {
782 rbi->bi_next = return_bi;
783 return_bi = rbi;
784 }
785 rbi = rbi2;
786 }
787 }
788 }
789 spin_unlock_irq(&conf->device_lock);
790 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
791
792 return_io(return_bi);
793
794 set_bit(STRIPE_HANDLE, &sh->state);
795 release_stripe(sh);
796}
797
798static void ops_run_biofill(struct stripe_head *sh)
799{
800 struct dma_async_tx_descriptor *tx = NULL;
801 struct r5conf *conf = sh->raid_conf;
802 struct async_submit_ctl submit;
803 int i;
804
805 pr_debug("%s: stripe %llu\n", __func__,
806 (unsigned long long)sh->sector);
807
808 for (i = sh->disks; i--; ) {
809 struct r5dev *dev = &sh->dev[i];
810 if (test_bit(R5_Wantfill, &dev->flags)) {
811 struct bio *rbi;
812 spin_lock_irq(&conf->device_lock);
813 dev->read = rbi = dev->toread;
814 dev->toread = NULL;
815 spin_unlock_irq(&conf->device_lock);
816 while (rbi && rbi->bi_sector <
817 dev->sector + STRIPE_SECTORS) {
818 tx = async_copy_data(0, rbi, dev->page,
819 dev->sector, tx);
820 rbi = r5_next_bio(rbi, dev->sector);
821 }
822 }
823 }
824
825 atomic_inc(&sh->count);
826 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
827 async_trigger_callback(&submit);
828}
829
830static void mark_target_uptodate(struct stripe_head *sh, int target)
831{
832 struct r5dev *tgt;
833
834 if (target < 0)
835 return;
836
837 tgt = &sh->dev[target];
838 set_bit(R5_UPTODATE, &tgt->flags);
839 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
840 clear_bit(R5_Wantcompute, &tgt->flags);
841}
842
843static void ops_complete_compute(void *stripe_head_ref)
844{
845 struct stripe_head *sh = stripe_head_ref;
846
847 pr_debug("%s: stripe %llu\n", __func__,
848 (unsigned long long)sh->sector);
849
850 /* mark the computed target(s) as uptodate */
851 mark_target_uptodate(sh, sh->ops.target);
852 mark_target_uptodate(sh, sh->ops.target2);
853
854 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
855 if (sh->check_state == check_state_compute_run)
856 sh->check_state = check_state_compute_result;
857 set_bit(STRIPE_HANDLE, &sh->state);
858 release_stripe(sh);
859}
860
861/* return a pointer to the address conversion region of the scribble buffer */
862static addr_conv_t *to_addr_conv(struct stripe_head *sh,
863 struct raid5_percpu *percpu)
864{
865 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
866}
867
868static struct dma_async_tx_descriptor *
869ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
870{
871 int disks = sh->disks;
872 struct page **xor_srcs = percpu->scribble;
873 int target = sh->ops.target;
874 struct r5dev *tgt = &sh->dev[target];
875 struct page *xor_dest = tgt->page;
876 int count = 0;
877 struct dma_async_tx_descriptor *tx;
878 struct async_submit_ctl submit;
879 int i;
880
881 pr_debug("%s: stripe %llu block: %d\n",
882 __func__, (unsigned long long)sh->sector, target);
883 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
884
885 for (i = disks; i--; )
886 if (i != target)
887 xor_srcs[count++] = sh->dev[i].page;
888
889 atomic_inc(&sh->count);
890
891 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
892 ops_complete_compute, sh, to_addr_conv(sh, percpu));
893 if (unlikely(count == 1))
894 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
895 else
896 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
897
898 return tx;
899}
900
901/* set_syndrome_sources - populate source buffers for gen_syndrome
902 * @srcs - (struct page *) array of size sh->disks
903 * @sh - stripe_head to parse
904 *
905 * Populates srcs in proper layout order for the stripe and returns the
906 * 'count' of sources to be used in a call to async_gen_syndrome. The P
907 * destination buffer is recorded in srcs[count] and the Q destination
908 * is recorded in srcs[count+1]].
909 */
910static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
911{
912 int disks = sh->disks;
913 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
914 int d0_idx = raid6_d0(sh);
915 int count;
916 int i;
917
918 for (i = 0; i < disks; i++)
919 srcs[i] = NULL;
920
921 count = 0;
922 i = d0_idx;
923 do {
924 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
925
926 srcs[slot] = sh->dev[i].page;
927 i = raid6_next_disk(i, disks);
928 } while (i != d0_idx);
929
930 return syndrome_disks;
931}
932
933static struct dma_async_tx_descriptor *
934ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
935{
936 int disks = sh->disks;
937 struct page **blocks = percpu->scribble;
938 int target;
939 int qd_idx = sh->qd_idx;
940 struct dma_async_tx_descriptor *tx;
941 struct async_submit_ctl submit;
942 struct r5dev *tgt;
943 struct page *dest;
944 int i;
945 int count;
946
947 if (sh->ops.target < 0)
948 target = sh->ops.target2;
949 else if (sh->ops.target2 < 0)
950 target = sh->ops.target;
951 else
952 /* we should only have one valid target */
953 BUG();
954 BUG_ON(target < 0);
955 pr_debug("%s: stripe %llu block: %d\n",
956 __func__, (unsigned long long)sh->sector, target);
957
958 tgt = &sh->dev[target];
959 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
960 dest = tgt->page;
961
962 atomic_inc(&sh->count);
963
964 if (target == qd_idx) {
965 count = set_syndrome_sources(blocks, sh);
966 blocks[count] = NULL; /* regenerating p is not necessary */
967 BUG_ON(blocks[count+1] != dest); /* q should already be set */
968 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
969 ops_complete_compute, sh,
970 to_addr_conv(sh, percpu));
971 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
972 } else {
973 /* Compute any data- or p-drive using XOR */
974 count = 0;
975 for (i = disks; i-- ; ) {
976 if (i == target || i == qd_idx)
977 continue;
978 blocks[count++] = sh->dev[i].page;
979 }
980
981 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
982 NULL, ops_complete_compute, sh,
983 to_addr_conv(sh, percpu));
984 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
985 }
986
987 return tx;
988}
989
990static struct dma_async_tx_descriptor *
991ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
992{
993 int i, count, disks = sh->disks;
994 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
995 int d0_idx = raid6_d0(sh);
996 int faila = -1, failb = -1;
997 int target = sh->ops.target;
998 int target2 = sh->ops.target2;
999 struct r5dev *tgt = &sh->dev[target];
1000 struct r5dev *tgt2 = &sh->dev[target2];
1001 struct dma_async_tx_descriptor *tx;
1002 struct page **blocks = percpu->scribble;
1003 struct async_submit_ctl submit;
1004
1005 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1006 __func__, (unsigned long long)sh->sector, target, target2);
1007 BUG_ON(target < 0 || target2 < 0);
1008 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1009 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1010
1011 /* we need to open-code set_syndrome_sources to handle the
1012 * slot number conversion for 'faila' and 'failb'
1013 */
1014 for (i = 0; i < disks ; i++)
1015 blocks[i] = NULL;
1016 count = 0;
1017 i = d0_idx;
1018 do {
1019 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1020
1021 blocks[slot] = sh->dev[i].page;
1022
1023 if (i == target)
1024 faila = slot;
1025 if (i == target2)
1026 failb = slot;
1027 i = raid6_next_disk(i, disks);
1028 } while (i != d0_idx);
1029
1030 BUG_ON(faila == failb);
1031 if (failb < faila)
1032 swap(faila, failb);
1033 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1034 __func__, (unsigned long long)sh->sector, faila, failb);
1035
1036 atomic_inc(&sh->count);
1037
1038 if (failb == syndrome_disks+1) {
1039 /* Q disk is one of the missing disks */
1040 if (faila == syndrome_disks) {
1041 /* Missing P+Q, just recompute */
1042 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1043 ops_complete_compute, sh,
1044 to_addr_conv(sh, percpu));
1045 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1046 STRIPE_SIZE, &submit);
1047 } else {
1048 struct page *dest;
1049 int data_target;
1050 int qd_idx = sh->qd_idx;
1051
1052 /* Missing D+Q: recompute D from P, then recompute Q */
1053 if (target == qd_idx)
1054 data_target = target2;
1055 else
1056 data_target = target;
1057
1058 count = 0;
1059 for (i = disks; i-- ; ) {
1060 if (i == data_target || i == qd_idx)
1061 continue;
1062 blocks[count++] = sh->dev[i].page;
1063 }
1064 dest = sh->dev[data_target].page;
1065 init_async_submit(&submit,
1066 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1067 NULL, NULL, NULL,
1068 to_addr_conv(sh, percpu));
1069 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1070 &submit);
1071
1072 count = set_syndrome_sources(blocks, sh);
1073 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1074 ops_complete_compute, sh,
1075 to_addr_conv(sh, percpu));
1076 return async_gen_syndrome(blocks, 0, count+2,
1077 STRIPE_SIZE, &submit);
1078 }
1079 } else {
1080 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1081 ops_complete_compute, sh,
1082 to_addr_conv(sh, percpu));
1083 if (failb == syndrome_disks) {
1084 /* We're missing D+P. */
1085 return async_raid6_datap_recov(syndrome_disks+2,
1086 STRIPE_SIZE, faila,
1087 blocks, &submit);
1088 } else {
1089 /* We're missing D+D. */
1090 return async_raid6_2data_recov(syndrome_disks+2,
1091 STRIPE_SIZE, faila, failb,
1092 blocks, &submit);
1093 }
1094 }
1095}
1096
1097
1098static void ops_complete_prexor(void *stripe_head_ref)
1099{
1100 struct stripe_head *sh = stripe_head_ref;
1101
1102 pr_debug("%s: stripe %llu\n", __func__,
1103 (unsigned long long)sh->sector);
1104}
1105
1106static struct dma_async_tx_descriptor *
1107ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1108 struct dma_async_tx_descriptor *tx)
1109{
1110 int disks = sh->disks;
1111 struct page **xor_srcs = percpu->scribble;
1112 int count = 0, pd_idx = sh->pd_idx, i;
1113 struct async_submit_ctl submit;
1114
1115 /* existing parity data subtracted */
1116 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1117
1118 pr_debug("%s: stripe %llu\n", __func__,
1119 (unsigned long long)sh->sector);
1120
1121 for (i = disks; i--; ) {
1122 struct r5dev *dev = &sh->dev[i];
1123 /* Only process blocks that are known to be uptodate */
1124 if (test_bit(R5_Wantdrain, &dev->flags))
1125 xor_srcs[count++] = dev->page;
1126 }
1127
1128 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1129 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1130 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1131
1132 return tx;
1133}
1134
1135static struct dma_async_tx_descriptor *
1136ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1137{
1138 int disks = sh->disks;
1139 int i;
1140
1141 pr_debug("%s: stripe %llu\n", __func__,
1142 (unsigned long long)sh->sector);
1143
1144 for (i = disks; i--; ) {
1145 struct r5dev *dev = &sh->dev[i];
1146 struct bio *chosen;
1147
1148 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1149 struct bio *wbi;
1150
1151 spin_lock_irq(&sh->raid_conf->device_lock);
1152 chosen = dev->towrite;
1153 dev->towrite = NULL;
1154 BUG_ON(dev->written);
1155 wbi = dev->written = chosen;
1156 spin_unlock_irq(&sh->raid_conf->device_lock);
1157
1158 while (wbi && wbi->bi_sector <
1159 dev->sector + STRIPE_SECTORS) {
1160 if (wbi->bi_rw & REQ_FUA)
1161 set_bit(R5_WantFUA, &dev->flags);
1162 if (wbi->bi_rw & REQ_SYNC)
1163 set_bit(R5_SyncIO, &dev->flags);
1164 tx = async_copy_data(1, wbi, dev->page,
1165 dev->sector, tx);
1166 wbi = r5_next_bio(wbi, dev->sector);
1167 }
1168 }
1169 }
1170
1171 return tx;
1172}
1173
1174static void ops_complete_reconstruct(void *stripe_head_ref)
1175{
1176 struct stripe_head *sh = stripe_head_ref;
1177 int disks = sh->disks;
1178 int pd_idx = sh->pd_idx;
1179 int qd_idx = sh->qd_idx;
1180 int i;
1181 bool fua = false, sync = false;
1182
1183 pr_debug("%s: stripe %llu\n", __func__,
1184 (unsigned long long)sh->sector);
1185
1186 for (i = disks; i--; ) {
1187 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1188 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1189 }
1190
1191 for (i = disks; i--; ) {
1192 struct r5dev *dev = &sh->dev[i];
1193
1194 if (dev->written || i == pd_idx || i == qd_idx) {
1195 set_bit(R5_UPTODATE, &dev->flags);
1196 if (fua)
1197 set_bit(R5_WantFUA, &dev->flags);
1198 if (sync)
1199 set_bit(R5_SyncIO, &dev->flags);
1200 }
1201 }
1202
1203 if (sh->reconstruct_state == reconstruct_state_drain_run)
1204 sh->reconstruct_state = reconstruct_state_drain_result;
1205 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1206 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1207 else {
1208 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1209 sh->reconstruct_state = reconstruct_state_result;
1210 }
1211
1212 set_bit(STRIPE_HANDLE, &sh->state);
1213 release_stripe(sh);
1214}
1215
1216static void
1217ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1218 struct dma_async_tx_descriptor *tx)
1219{
1220 int disks = sh->disks;
1221 struct page **xor_srcs = percpu->scribble;
1222 struct async_submit_ctl submit;
1223 int count = 0, pd_idx = sh->pd_idx, i;
1224 struct page *xor_dest;
1225 int prexor = 0;
1226 unsigned long flags;
1227
1228 pr_debug("%s: stripe %llu\n", __func__,
1229 (unsigned long long)sh->sector);
1230
1231 /* check if prexor is active which means only process blocks
1232 * that are part of a read-modify-write (written)
1233 */
1234 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1235 prexor = 1;
1236 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1237 for (i = disks; i--; ) {
1238 struct r5dev *dev = &sh->dev[i];
1239 if (dev->written)
1240 xor_srcs[count++] = dev->page;
1241 }
1242 } else {
1243 xor_dest = sh->dev[pd_idx].page;
1244 for (i = disks; i--; ) {
1245 struct r5dev *dev = &sh->dev[i];
1246 if (i != pd_idx)
1247 xor_srcs[count++] = dev->page;
1248 }
1249 }
1250
1251 /* 1/ if we prexor'd then the dest is reused as a source
1252 * 2/ if we did not prexor then we are redoing the parity
1253 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1254 * for the synchronous xor case
1255 */
1256 flags = ASYNC_TX_ACK |
1257 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1258
1259 atomic_inc(&sh->count);
1260
1261 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1262 to_addr_conv(sh, percpu));
1263 if (unlikely(count == 1))
1264 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1265 else
1266 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1267}
1268
1269static void
1270ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1271 struct dma_async_tx_descriptor *tx)
1272{
1273 struct async_submit_ctl submit;
1274 struct page **blocks = percpu->scribble;
1275 int count;
1276
1277 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1278
1279 count = set_syndrome_sources(blocks, sh);
1280
1281 atomic_inc(&sh->count);
1282
1283 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1284 sh, to_addr_conv(sh, percpu));
1285 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1286}
1287
1288static void ops_complete_check(void *stripe_head_ref)
1289{
1290 struct stripe_head *sh = stripe_head_ref;
1291
1292 pr_debug("%s: stripe %llu\n", __func__,
1293 (unsigned long long)sh->sector);
1294
1295 sh->check_state = check_state_check_result;
1296 set_bit(STRIPE_HANDLE, &sh->state);
1297 release_stripe(sh);
1298}
1299
1300static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1301{
1302 int disks = sh->disks;
1303 int pd_idx = sh->pd_idx;
1304 int qd_idx = sh->qd_idx;
1305 struct page *xor_dest;
1306 struct page **xor_srcs = percpu->scribble;
1307 struct dma_async_tx_descriptor *tx;
1308 struct async_submit_ctl submit;
1309 int count;
1310 int i;
1311
1312 pr_debug("%s: stripe %llu\n", __func__,
1313 (unsigned long long)sh->sector);
1314
1315 count = 0;
1316 xor_dest = sh->dev[pd_idx].page;
1317 xor_srcs[count++] = xor_dest;
1318 for (i = disks; i--; ) {
1319 if (i == pd_idx || i == qd_idx)
1320 continue;
1321 xor_srcs[count++] = sh->dev[i].page;
1322 }
1323
1324 init_async_submit(&submit, 0, NULL, NULL, NULL,
1325 to_addr_conv(sh, percpu));
1326 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1327 &sh->ops.zero_sum_result, &submit);
1328
1329 atomic_inc(&sh->count);
1330 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1331 tx = async_trigger_callback(&submit);
1332}
1333
1334static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1335{
1336 struct page **srcs = percpu->scribble;
1337 struct async_submit_ctl submit;
1338 int count;
1339
1340 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1341 (unsigned long long)sh->sector, checkp);
1342
1343 count = set_syndrome_sources(srcs, sh);
1344 if (!checkp)
1345 srcs[count] = NULL;
1346
1347 atomic_inc(&sh->count);
1348 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1349 sh, to_addr_conv(sh, percpu));
1350 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1351 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1352}
1353
1354static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1355{
1356 int overlap_clear = 0, i, disks = sh->disks;
1357 struct dma_async_tx_descriptor *tx = NULL;
1358 struct r5conf *conf = sh->raid_conf;
1359 int level = conf->level;
1360 struct raid5_percpu *percpu;
1361 unsigned long cpu;
1362
1363 cpu = get_cpu();
1364 percpu = per_cpu_ptr(conf->percpu, cpu);
1365 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1366 ops_run_biofill(sh);
1367 overlap_clear++;
1368 }
1369
1370 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1371 if (level < 6)
1372 tx = ops_run_compute5(sh, percpu);
1373 else {
1374 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1375 tx = ops_run_compute6_1(sh, percpu);
1376 else
1377 tx = ops_run_compute6_2(sh, percpu);
1378 }
1379 /* terminate the chain if reconstruct is not set to be run */
1380 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1381 async_tx_ack(tx);
1382 }
1383
1384 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1385 tx = ops_run_prexor(sh, percpu, tx);
1386
1387 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1388 tx = ops_run_biodrain(sh, tx);
1389 overlap_clear++;
1390 }
1391
1392 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1393 if (level < 6)
1394 ops_run_reconstruct5(sh, percpu, tx);
1395 else
1396 ops_run_reconstruct6(sh, percpu, tx);
1397 }
1398
1399 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1400 if (sh->check_state == check_state_run)
1401 ops_run_check_p(sh, percpu);
1402 else if (sh->check_state == check_state_run_q)
1403 ops_run_check_pq(sh, percpu, 0);
1404 else if (sh->check_state == check_state_run_pq)
1405 ops_run_check_pq(sh, percpu, 1);
1406 else
1407 BUG();
1408 }
1409
1410 if (overlap_clear)
1411 for (i = disks; i--; ) {
1412 struct r5dev *dev = &sh->dev[i];
1413 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1414 wake_up(&sh->raid_conf->wait_for_overlap);
1415 }
1416 put_cpu();
1417}
1418
1419#ifdef CONFIG_MULTICORE_RAID456
1420static void async_run_ops(void *param, async_cookie_t cookie)
1421{
1422 struct stripe_head *sh = param;
1423 unsigned long ops_request = sh->ops.request;
1424
1425 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1426 wake_up(&sh->ops.wait_for_ops);
1427
1428 __raid_run_ops(sh, ops_request);
1429 release_stripe(sh);
1430}
1431
1432static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1433{
1434 /* since handle_stripe can be called outside of raid5d context
1435 * we need to ensure sh->ops.request is de-staged before another
1436 * request arrives
1437 */
1438 wait_event(sh->ops.wait_for_ops,
1439 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1440 sh->ops.request = ops_request;
1441
1442 atomic_inc(&sh->count);
1443 async_schedule(async_run_ops, sh);
1444}
1445#else
1446#define raid_run_ops __raid_run_ops
1447#endif
1448
1449static int grow_one_stripe(struct r5conf *conf)
1450{
1451 struct stripe_head *sh;
1452 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1453 if (!sh)
1454 return 0;
1455
1456 sh->raid_conf = conf;
1457 #ifdef CONFIG_MULTICORE_RAID456
1458 init_waitqueue_head(&sh->ops.wait_for_ops);
1459 #endif
1460
1461 if (grow_buffers(sh)) {
1462 shrink_buffers(sh);
1463 kmem_cache_free(conf->slab_cache, sh);
1464 return 0;
1465 }
1466 /* we just created an active stripe so... */
1467 atomic_set(&sh->count, 1);
1468 atomic_inc(&conf->active_stripes);
1469 INIT_LIST_HEAD(&sh->lru);
1470 release_stripe(sh);
1471 return 1;
1472}
1473
1474static int grow_stripes(struct r5conf *conf, int num)
1475{
1476 struct kmem_cache *sc;
1477 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1478
1479 if (conf->mddev->gendisk)
1480 sprintf(conf->cache_name[0],
1481 "raid%d-%s", conf->level, mdname(conf->mddev));
1482 else
1483 sprintf(conf->cache_name[0],
1484 "raid%d-%p", conf->level, conf->mddev);
1485 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1486
1487 conf->active_name = 0;
1488 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1489 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1490 0, 0, NULL);
1491 if (!sc)
1492 return 1;
1493 conf->slab_cache = sc;
1494 conf->pool_size = devs;
1495 while (num--)
1496 if (!grow_one_stripe(conf))
1497 return 1;
1498 return 0;
1499}
1500
1501/**
1502 * scribble_len - return the required size of the scribble region
1503 * @num - total number of disks in the array
1504 *
1505 * The size must be enough to contain:
1506 * 1/ a struct page pointer for each device in the array +2
1507 * 2/ room to convert each entry in (1) to its corresponding dma
1508 * (dma_map_page()) or page (page_address()) address.
1509 *
1510 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1511 * calculate over all devices (not just the data blocks), using zeros in place
1512 * of the P and Q blocks.
1513 */
1514static size_t scribble_len(int num)
1515{
1516 size_t len;
1517
1518 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1519
1520 return len;
1521}
1522
1523static int resize_stripes(struct r5conf *conf, int newsize)
1524{
1525 /* Make all the stripes able to hold 'newsize' devices.
1526 * New slots in each stripe get 'page' set to a new page.
1527 *
1528 * This happens in stages:
1529 * 1/ create a new kmem_cache and allocate the required number of
1530 * stripe_heads.
1531 * 2/ gather all the old stripe_heads and tranfer the pages across
1532 * to the new stripe_heads. This will have the side effect of
1533 * freezing the array as once all stripe_heads have been collected,
1534 * no IO will be possible. Old stripe heads are freed once their
1535 * pages have been transferred over, and the old kmem_cache is
1536 * freed when all stripes are done.
1537 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1538 * we simple return a failre status - no need to clean anything up.
1539 * 4/ allocate new pages for the new slots in the new stripe_heads.
1540 * If this fails, we don't bother trying the shrink the
1541 * stripe_heads down again, we just leave them as they are.
1542 * As each stripe_head is processed the new one is released into
1543 * active service.
1544 *
1545 * Once step2 is started, we cannot afford to wait for a write,
1546 * so we use GFP_NOIO allocations.
1547 */
1548 struct stripe_head *osh, *nsh;
1549 LIST_HEAD(newstripes);
1550 struct disk_info *ndisks;
1551 unsigned long cpu;
1552 int err;
1553 struct kmem_cache *sc;
1554 int i;
1555
1556 if (newsize <= conf->pool_size)
1557 return 0; /* never bother to shrink */
1558
1559 err = md_allow_write(conf->mddev);
1560 if (err)
1561 return err;
1562
1563 /* Step 1 */
1564 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1565 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1566 0, 0, NULL);
1567 if (!sc)
1568 return -ENOMEM;
1569
1570 for (i = conf->max_nr_stripes; i; i--) {
1571 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1572 if (!nsh)
1573 break;
1574
1575 nsh->raid_conf = conf;
1576 #ifdef CONFIG_MULTICORE_RAID456
1577 init_waitqueue_head(&nsh->ops.wait_for_ops);
1578 #endif
1579
1580 list_add(&nsh->lru, &newstripes);
1581 }
1582 if (i) {
1583 /* didn't get enough, give up */
1584 while (!list_empty(&newstripes)) {
1585 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1586 list_del(&nsh->lru);
1587 kmem_cache_free(sc, nsh);
1588 }
1589 kmem_cache_destroy(sc);
1590 return -ENOMEM;
1591 }
1592 /* Step 2 - Must use GFP_NOIO now.
1593 * OK, we have enough stripes, start collecting inactive
1594 * stripes and copying them over
1595 */
1596 list_for_each_entry(nsh, &newstripes, lru) {
1597 spin_lock_irq(&conf->device_lock);
1598 wait_event_lock_irq(conf->wait_for_stripe,
1599 !list_empty(&conf->inactive_list),
1600 conf->device_lock,
1601 );
1602 osh = get_free_stripe(conf);
1603 spin_unlock_irq(&conf->device_lock);
1604 atomic_set(&nsh->count, 1);
1605 for(i=0; i<conf->pool_size; i++)
1606 nsh->dev[i].page = osh->dev[i].page;
1607 for( ; i<newsize; i++)
1608 nsh->dev[i].page = NULL;
1609 kmem_cache_free(conf->slab_cache, osh);
1610 }
1611 kmem_cache_destroy(conf->slab_cache);
1612
1613 /* Step 3.
1614 * At this point, we are holding all the stripes so the array
1615 * is completely stalled, so now is a good time to resize
1616 * conf->disks and the scribble region
1617 */
1618 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1619 if (ndisks) {
1620 for (i=0; i<conf->raid_disks; i++)
1621 ndisks[i] = conf->disks[i];
1622 kfree(conf->disks);
1623 conf->disks = ndisks;
1624 } else
1625 err = -ENOMEM;
1626
1627 get_online_cpus();
1628 conf->scribble_len = scribble_len(newsize);
1629 for_each_present_cpu(cpu) {
1630 struct raid5_percpu *percpu;
1631 void *scribble;
1632
1633 percpu = per_cpu_ptr(conf->percpu, cpu);
1634 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1635
1636 if (scribble) {
1637 kfree(percpu->scribble);
1638 percpu->scribble = scribble;
1639 } else {
1640 err = -ENOMEM;
1641 break;
1642 }
1643 }
1644 put_online_cpus();
1645
1646 /* Step 4, return new stripes to service */
1647 while(!list_empty(&newstripes)) {
1648 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1649 list_del_init(&nsh->lru);
1650
1651 for (i=conf->raid_disks; i < newsize; i++)
1652 if (nsh->dev[i].page == NULL) {
1653 struct page *p = alloc_page(GFP_NOIO);
1654 nsh->dev[i].page = p;
1655 if (!p)
1656 err = -ENOMEM;
1657 }
1658 release_stripe(nsh);
1659 }
1660 /* critical section pass, GFP_NOIO no longer needed */
1661
1662 conf->slab_cache = sc;
1663 conf->active_name = 1-conf->active_name;
1664 conf->pool_size = newsize;
1665 return err;
1666}
1667
1668static int drop_one_stripe(struct r5conf *conf)
1669{
1670 struct stripe_head *sh;
1671
1672 spin_lock_irq(&conf->device_lock);
1673 sh = get_free_stripe(conf);
1674 spin_unlock_irq(&conf->device_lock);
1675 if (!sh)
1676 return 0;
1677 BUG_ON(atomic_read(&sh->count));
1678 shrink_buffers(sh);
1679 kmem_cache_free(conf->slab_cache, sh);
1680 atomic_dec(&conf->active_stripes);
1681 return 1;
1682}
1683
1684static void shrink_stripes(struct r5conf *conf)
1685{
1686 while (drop_one_stripe(conf))
1687 ;
1688
1689 if (conf->slab_cache)
1690 kmem_cache_destroy(conf->slab_cache);
1691 conf->slab_cache = NULL;
1692}
1693
1694static void raid5_end_read_request(struct bio * bi, int error)
1695{
1696 struct stripe_head *sh = bi->bi_private;
1697 struct r5conf *conf = sh->raid_conf;
1698 int disks = sh->disks, i;
1699 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1700 char b[BDEVNAME_SIZE];
1701 struct md_rdev *rdev = NULL;
1702 sector_t s;
1703
1704 for (i=0 ; i<disks; i++)
1705 if (bi == &sh->dev[i].req)
1706 break;
1707
1708 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1709 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1710 uptodate);
1711 if (i == disks) {
1712 BUG();
1713 return;
1714 }
1715 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1716 /* If replacement finished while this request was outstanding,
1717 * 'replacement' might be NULL already.
1718 * In that case it moved down to 'rdev'.
1719 * rdev is not removed until all requests are finished.
1720 */
1721 rdev = conf->disks[i].replacement;
1722 if (!rdev)
1723 rdev = conf->disks[i].rdev;
1724
1725 if (use_new_offset(conf, sh))
1726 s = sh->sector + rdev->new_data_offset;
1727 else
1728 s = sh->sector + rdev->data_offset;
1729 if (uptodate) {
1730 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1731 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1732 /* Note that this cannot happen on a
1733 * replacement device. We just fail those on
1734 * any error
1735 */
1736 printk_ratelimited(
1737 KERN_INFO
1738 "md/raid:%s: read error corrected"
1739 " (%lu sectors at %llu on %s)\n",
1740 mdname(conf->mddev), STRIPE_SECTORS,
1741 (unsigned long long)s,
1742 bdevname(rdev->bdev, b));
1743 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1744 clear_bit(R5_ReadError, &sh->dev[i].flags);
1745 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1746 }
1747 if (atomic_read(&rdev->read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 } else {
1750 const char *bdn = bdevname(rdev->bdev, b);
1751 int retry = 0;
1752 int set_bad = 0;
1753
1754 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1755 atomic_inc(&rdev->read_errors);
1756 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1757 printk_ratelimited(
1758 KERN_WARNING
1759 "md/raid:%s: read error on replacement device "
1760 "(sector %llu on %s).\n",
1761 mdname(conf->mddev),
1762 (unsigned long long)s,
1763 bdn);
1764 else if (conf->mddev->degraded >= conf->max_degraded) {
1765 set_bad = 1;
1766 printk_ratelimited(
1767 KERN_WARNING
1768 "md/raid:%s: read error not correctable "
1769 "(sector %llu on %s).\n",
1770 mdname(conf->mddev),
1771 (unsigned long long)s,
1772 bdn);
1773 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1774 /* Oh, no!!! */
1775 set_bad = 1;
1776 printk_ratelimited(
1777 KERN_WARNING
1778 "md/raid:%s: read error NOT corrected!! "
1779 "(sector %llu on %s).\n",
1780 mdname(conf->mddev),
1781 (unsigned long long)s,
1782 bdn);
1783 } else if (atomic_read(&rdev->read_errors)
1784 > conf->max_nr_stripes)
1785 printk(KERN_WARNING
1786 "md/raid:%s: Too many read errors, failing device %s.\n",
1787 mdname(conf->mddev), bdn);
1788 else
1789 retry = 1;
1790 if (retry)
1791 set_bit(R5_ReadError, &sh->dev[i].flags);
1792 else {
1793 clear_bit(R5_ReadError, &sh->dev[i].flags);
1794 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1795 if (!(set_bad
1796 && test_bit(In_sync, &rdev->flags)
1797 && rdev_set_badblocks(
1798 rdev, sh->sector, STRIPE_SECTORS, 0)))
1799 md_error(conf->mddev, rdev);
1800 }
1801 }
1802 rdev_dec_pending(rdev, conf->mddev);
1803 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1804 set_bit(STRIPE_HANDLE, &sh->state);
1805 release_stripe(sh);
1806}
1807
1808static void raid5_end_write_request(struct bio *bi, int error)
1809{
1810 struct stripe_head *sh = bi->bi_private;
1811 struct r5conf *conf = sh->raid_conf;
1812 int disks = sh->disks, i;
1813 struct md_rdev *uninitialized_var(rdev);
1814 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1815 sector_t first_bad;
1816 int bad_sectors;
1817 int replacement = 0;
1818
1819 for (i = 0 ; i < disks; i++) {
1820 if (bi == &sh->dev[i].req) {
1821 rdev = conf->disks[i].rdev;
1822 break;
1823 }
1824 if (bi == &sh->dev[i].rreq) {
1825 rdev = conf->disks[i].replacement;
1826 if (rdev)
1827 replacement = 1;
1828 else
1829 /* rdev was removed and 'replacement'
1830 * replaced it. rdev is not removed
1831 * until all requests are finished.
1832 */
1833 rdev = conf->disks[i].rdev;
1834 break;
1835 }
1836 }
1837 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1838 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1839 uptodate);
1840 if (i == disks) {
1841 BUG();
1842 return;
1843 }
1844
1845 if (replacement) {
1846 if (!uptodate)
1847 md_error(conf->mddev, rdev);
1848 else if (is_badblock(rdev, sh->sector,
1849 STRIPE_SECTORS,
1850 &first_bad, &bad_sectors))
1851 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1852 } else {
1853 if (!uptodate) {
1854 set_bit(WriteErrorSeen, &rdev->flags);
1855 set_bit(R5_WriteError, &sh->dev[i].flags);
1856 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1857 set_bit(MD_RECOVERY_NEEDED,
1858 &rdev->mddev->recovery);
1859 } else if (is_badblock(rdev, sh->sector,
1860 STRIPE_SECTORS,
1861 &first_bad, &bad_sectors))
1862 set_bit(R5_MadeGood, &sh->dev[i].flags);
1863 }
1864 rdev_dec_pending(rdev, conf->mddev);
1865
1866 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1867 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1868 set_bit(STRIPE_HANDLE, &sh->state);
1869 release_stripe(sh);
1870}
1871
1872static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1873
1874static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1875{
1876 struct r5dev *dev = &sh->dev[i];
1877
1878 bio_init(&dev->req);
1879 dev->req.bi_io_vec = &dev->vec;
1880 dev->req.bi_vcnt++;
1881 dev->req.bi_max_vecs++;
1882 dev->req.bi_private = sh;
1883 dev->vec.bv_page = dev->page;
1884
1885 bio_init(&dev->rreq);
1886 dev->rreq.bi_io_vec = &dev->rvec;
1887 dev->rreq.bi_vcnt++;
1888 dev->rreq.bi_max_vecs++;
1889 dev->rreq.bi_private = sh;
1890 dev->rvec.bv_page = dev->page;
1891
1892 dev->flags = 0;
1893 dev->sector = compute_blocknr(sh, i, previous);
1894}
1895
1896static void error(struct mddev *mddev, struct md_rdev *rdev)
1897{
1898 char b[BDEVNAME_SIZE];
1899 struct r5conf *conf = mddev->private;
1900 unsigned long flags;
1901 pr_debug("raid456: error called\n");
1902
1903 spin_lock_irqsave(&conf->device_lock, flags);
1904 clear_bit(In_sync, &rdev->flags);
1905 mddev->degraded = calc_degraded(conf);
1906 spin_unlock_irqrestore(&conf->device_lock, flags);
1907 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1908
1909 set_bit(Blocked, &rdev->flags);
1910 set_bit(Faulty, &rdev->flags);
1911 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1912 printk(KERN_ALERT
1913 "md/raid:%s: Disk failure on %s, disabling device.\n"
1914 "md/raid:%s: Operation continuing on %d devices.\n",
1915 mdname(mddev),
1916 bdevname(rdev->bdev, b),
1917 mdname(mddev),
1918 conf->raid_disks - mddev->degraded);
1919}
1920
1921/*
1922 * Input: a 'big' sector number,
1923 * Output: index of the data and parity disk, and the sector # in them.
1924 */
1925static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1926 int previous, int *dd_idx,
1927 struct stripe_head *sh)
1928{
1929 sector_t stripe, stripe2;
1930 sector_t chunk_number;
1931 unsigned int chunk_offset;
1932 int pd_idx, qd_idx;
1933 int ddf_layout = 0;
1934 sector_t new_sector;
1935 int algorithm = previous ? conf->prev_algo
1936 : conf->algorithm;
1937 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1938 : conf->chunk_sectors;
1939 int raid_disks = previous ? conf->previous_raid_disks
1940 : conf->raid_disks;
1941 int data_disks = raid_disks - conf->max_degraded;
1942
1943 /* First compute the information on this sector */
1944
1945 /*
1946 * Compute the chunk number and the sector offset inside the chunk
1947 */
1948 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1949 chunk_number = r_sector;
1950
1951 /*
1952 * Compute the stripe number
1953 */
1954 stripe = chunk_number;
1955 *dd_idx = sector_div(stripe, data_disks);
1956 stripe2 = stripe;
1957 /*
1958 * Select the parity disk based on the user selected algorithm.
1959 */
1960 pd_idx = qd_idx = -1;
1961 switch(conf->level) {
1962 case 4:
1963 pd_idx = data_disks;
1964 break;
1965 case 5:
1966 switch (algorithm) {
1967 case ALGORITHM_LEFT_ASYMMETRIC:
1968 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1969 if (*dd_idx >= pd_idx)
1970 (*dd_idx)++;
1971 break;
1972 case ALGORITHM_RIGHT_ASYMMETRIC:
1973 pd_idx = sector_div(stripe2, raid_disks);
1974 if (*dd_idx >= pd_idx)
1975 (*dd_idx)++;
1976 break;
1977 case ALGORITHM_LEFT_SYMMETRIC:
1978 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1979 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1980 break;
1981 case ALGORITHM_RIGHT_SYMMETRIC:
1982 pd_idx = sector_div(stripe2, raid_disks);
1983 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1984 break;
1985 case ALGORITHM_PARITY_0:
1986 pd_idx = 0;
1987 (*dd_idx)++;
1988 break;
1989 case ALGORITHM_PARITY_N:
1990 pd_idx = data_disks;
1991 break;
1992 default:
1993 BUG();
1994 }
1995 break;
1996 case 6:
1997
1998 switch (algorithm) {
1999 case ALGORITHM_LEFT_ASYMMETRIC:
2000 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2001 qd_idx = pd_idx + 1;
2002 if (pd_idx == raid_disks-1) {
2003 (*dd_idx)++; /* Q D D D P */
2004 qd_idx = 0;
2005 } else if (*dd_idx >= pd_idx)
2006 (*dd_idx) += 2; /* D D P Q D */
2007 break;
2008 case ALGORITHM_RIGHT_ASYMMETRIC:
2009 pd_idx = sector_div(stripe2, raid_disks);
2010 qd_idx = pd_idx + 1;
2011 if (pd_idx == raid_disks-1) {
2012 (*dd_idx)++; /* Q D D D P */
2013 qd_idx = 0;
2014 } else if (*dd_idx >= pd_idx)
2015 (*dd_idx) += 2; /* D D P Q D */
2016 break;
2017 case ALGORITHM_LEFT_SYMMETRIC:
2018 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2019 qd_idx = (pd_idx + 1) % raid_disks;
2020 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2021 break;
2022 case ALGORITHM_RIGHT_SYMMETRIC:
2023 pd_idx = sector_div(stripe2, raid_disks);
2024 qd_idx = (pd_idx + 1) % raid_disks;
2025 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2026 break;
2027
2028 case ALGORITHM_PARITY_0:
2029 pd_idx = 0;
2030 qd_idx = 1;
2031 (*dd_idx) += 2;
2032 break;
2033 case ALGORITHM_PARITY_N:
2034 pd_idx = data_disks;
2035 qd_idx = data_disks + 1;
2036 break;
2037
2038 case ALGORITHM_ROTATING_ZERO_RESTART:
2039 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2040 * of blocks for computing Q is different.
2041 */
2042 pd_idx = sector_div(stripe2, raid_disks);
2043 qd_idx = pd_idx + 1;
2044 if (pd_idx == raid_disks-1) {
2045 (*dd_idx)++; /* Q D D D P */
2046 qd_idx = 0;
2047 } else if (*dd_idx >= pd_idx)
2048 (*dd_idx) += 2; /* D D P Q D */
2049 ddf_layout = 1;
2050 break;
2051
2052 case ALGORITHM_ROTATING_N_RESTART:
2053 /* Same a left_asymmetric, by first stripe is
2054 * D D D P Q rather than
2055 * Q D D D P
2056 */
2057 stripe2 += 1;
2058 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2059 qd_idx = pd_idx + 1;
2060 if (pd_idx == raid_disks-1) {
2061 (*dd_idx)++; /* Q D D D P */
2062 qd_idx = 0;
2063 } else if (*dd_idx >= pd_idx)
2064 (*dd_idx) += 2; /* D D P Q D */
2065 ddf_layout = 1;
2066 break;
2067
2068 case ALGORITHM_ROTATING_N_CONTINUE:
2069 /* Same as left_symmetric but Q is before P */
2070 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2071 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2072 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2073 ddf_layout = 1;
2074 break;
2075
2076 case ALGORITHM_LEFT_ASYMMETRIC_6:
2077 /* RAID5 left_asymmetric, with Q on last device */
2078 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2079 if (*dd_idx >= pd_idx)
2080 (*dd_idx)++;
2081 qd_idx = raid_disks - 1;
2082 break;
2083
2084 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2085 pd_idx = sector_div(stripe2, raid_disks-1);
2086 if (*dd_idx >= pd_idx)
2087 (*dd_idx)++;
2088 qd_idx = raid_disks - 1;
2089 break;
2090
2091 case ALGORITHM_LEFT_SYMMETRIC_6:
2092 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2093 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2094 qd_idx = raid_disks - 1;
2095 break;
2096
2097 case ALGORITHM_RIGHT_SYMMETRIC_6:
2098 pd_idx = sector_div(stripe2, raid_disks-1);
2099 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2100 qd_idx = raid_disks - 1;
2101 break;
2102
2103 case ALGORITHM_PARITY_0_6:
2104 pd_idx = 0;
2105 (*dd_idx)++;
2106 qd_idx = raid_disks - 1;
2107 break;
2108
2109 default:
2110 BUG();
2111 }
2112 break;
2113 }
2114
2115 if (sh) {
2116 sh->pd_idx = pd_idx;
2117 sh->qd_idx = qd_idx;
2118 sh->ddf_layout = ddf_layout;
2119 }
2120 /*
2121 * Finally, compute the new sector number
2122 */
2123 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2124 return new_sector;
2125}
2126
2127
2128static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2129{
2130 struct r5conf *conf = sh->raid_conf;
2131 int raid_disks = sh->disks;
2132 int data_disks = raid_disks - conf->max_degraded;
2133 sector_t new_sector = sh->sector, check;
2134 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2135 : conf->chunk_sectors;
2136 int algorithm = previous ? conf->prev_algo
2137 : conf->algorithm;
2138 sector_t stripe;
2139 int chunk_offset;
2140 sector_t chunk_number;
2141 int dummy1, dd_idx = i;
2142 sector_t r_sector;
2143 struct stripe_head sh2;
2144
2145
2146 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2147 stripe = new_sector;
2148
2149 if (i == sh->pd_idx)
2150 return 0;
2151 switch(conf->level) {
2152 case 4: break;
2153 case 5:
2154 switch (algorithm) {
2155 case ALGORITHM_LEFT_ASYMMETRIC:
2156 case ALGORITHM_RIGHT_ASYMMETRIC:
2157 if (i > sh->pd_idx)
2158 i--;
2159 break;
2160 case ALGORITHM_LEFT_SYMMETRIC:
2161 case ALGORITHM_RIGHT_SYMMETRIC:
2162 if (i < sh->pd_idx)
2163 i += raid_disks;
2164 i -= (sh->pd_idx + 1);
2165 break;
2166 case ALGORITHM_PARITY_0:
2167 i -= 1;
2168 break;
2169 case ALGORITHM_PARITY_N:
2170 break;
2171 default:
2172 BUG();
2173 }
2174 break;
2175 case 6:
2176 if (i == sh->qd_idx)
2177 return 0; /* It is the Q disk */
2178 switch (algorithm) {
2179 case ALGORITHM_LEFT_ASYMMETRIC:
2180 case ALGORITHM_RIGHT_ASYMMETRIC:
2181 case ALGORITHM_ROTATING_ZERO_RESTART:
2182 case ALGORITHM_ROTATING_N_RESTART:
2183 if (sh->pd_idx == raid_disks-1)
2184 i--; /* Q D D D P */
2185 else if (i > sh->pd_idx)
2186 i -= 2; /* D D P Q D */
2187 break;
2188 case ALGORITHM_LEFT_SYMMETRIC:
2189 case ALGORITHM_RIGHT_SYMMETRIC:
2190 if (sh->pd_idx == raid_disks-1)
2191 i--; /* Q D D D P */
2192 else {
2193 /* D D P Q D */
2194 if (i < sh->pd_idx)
2195 i += raid_disks;
2196 i -= (sh->pd_idx + 2);
2197 }
2198 break;
2199 case ALGORITHM_PARITY_0:
2200 i -= 2;
2201 break;
2202 case ALGORITHM_PARITY_N:
2203 break;
2204 case ALGORITHM_ROTATING_N_CONTINUE:
2205 /* Like left_symmetric, but P is before Q */
2206 if (sh->pd_idx == 0)
2207 i--; /* P D D D Q */
2208 else {
2209 /* D D Q P D */
2210 if (i < sh->pd_idx)
2211 i += raid_disks;
2212 i -= (sh->pd_idx + 1);
2213 }
2214 break;
2215 case ALGORITHM_LEFT_ASYMMETRIC_6:
2216 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2217 if (i > sh->pd_idx)
2218 i--;
2219 break;
2220 case ALGORITHM_LEFT_SYMMETRIC_6:
2221 case ALGORITHM_RIGHT_SYMMETRIC_6:
2222 if (i < sh->pd_idx)
2223 i += data_disks + 1;
2224 i -= (sh->pd_idx + 1);
2225 break;
2226 case ALGORITHM_PARITY_0_6:
2227 i -= 1;
2228 break;
2229 default:
2230 BUG();
2231 }
2232 break;
2233 }
2234
2235 chunk_number = stripe * data_disks + i;
2236 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2237
2238 check = raid5_compute_sector(conf, r_sector,
2239 previous, &dummy1, &sh2);
2240 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2241 || sh2.qd_idx != sh->qd_idx) {
2242 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2243 mdname(conf->mddev));
2244 return 0;
2245 }
2246 return r_sector;
2247}
2248
2249
2250static void
2251schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2252 int rcw, int expand)
2253{
2254 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2255 struct r5conf *conf = sh->raid_conf;
2256 int level = conf->level;
2257
2258 if (rcw) {
2259 /* if we are not expanding this is a proper write request, and
2260 * there will be bios with new data to be drained into the
2261 * stripe cache
2262 */
2263 if (!expand) {
2264 sh->reconstruct_state = reconstruct_state_drain_run;
2265 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2266 } else
2267 sh->reconstruct_state = reconstruct_state_run;
2268
2269 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2270
2271 for (i = disks; i--; ) {
2272 struct r5dev *dev = &sh->dev[i];
2273
2274 if (dev->towrite) {
2275 set_bit(R5_LOCKED, &dev->flags);
2276 set_bit(R5_Wantdrain, &dev->flags);
2277 if (!expand)
2278 clear_bit(R5_UPTODATE, &dev->flags);
2279 s->locked++;
2280 }
2281 }
2282 if (s->locked + conf->max_degraded == disks)
2283 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2284 atomic_inc(&conf->pending_full_writes);
2285 } else {
2286 BUG_ON(level == 6);
2287 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2288 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2289
2290 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2291 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2292 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2293 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2294
2295 for (i = disks; i--; ) {
2296 struct r5dev *dev = &sh->dev[i];
2297 if (i == pd_idx)
2298 continue;
2299
2300 if (dev->towrite &&
2301 (test_bit(R5_UPTODATE, &dev->flags) ||
2302 test_bit(R5_Wantcompute, &dev->flags))) {
2303 set_bit(R5_Wantdrain, &dev->flags);
2304 set_bit(R5_LOCKED, &dev->flags);
2305 clear_bit(R5_UPTODATE, &dev->flags);
2306 s->locked++;
2307 }
2308 }
2309 }
2310
2311 /* keep the parity disk(s) locked while asynchronous operations
2312 * are in flight
2313 */
2314 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2315 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2316 s->locked++;
2317
2318 if (level == 6) {
2319 int qd_idx = sh->qd_idx;
2320 struct r5dev *dev = &sh->dev[qd_idx];
2321
2322 set_bit(R5_LOCKED, &dev->flags);
2323 clear_bit(R5_UPTODATE, &dev->flags);
2324 s->locked++;
2325 }
2326
2327 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2328 __func__, (unsigned long long)sh->sector,
2329 s->locked, s->ops_request);
2330}
2331
2332/*
2333 * Each stripe/dev can have one or more bion attached.
2334 * toread/towrite point to the first in a chain.
2335 * The bi_next chain must be in order.
2336 */
2337static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2338{
2339 struct bio **bip;
2340 struct r5conf *conf = sh->raid_conf;
2341 int firstwrite=0;
2342
2343 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2344 (unsigned long long)bi->bi_sector,
2345 (unsigned long long)sh->sector);
2346
2347
2348 spin_lock_irq(&conf->device_lock);
2349 if (forwrite) {
2350 bip = &sh->dev[dd_idx].towrite;
2351 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2352 firstwrite = 1;
2353 } else
2354 bip = &sh->dev[dd_idx].toread;
2355 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2356 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2357 goto overlap;
2358 bip = & (*bip)->bi_next;
2359 }
2360 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2361 goto overlap;
2362
2363 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2364 if (*bip)
2365 bi->bi_next = *bip;
2366 *bip = bi;
2367 bi->bi_phys_segments++;
2368
2369 if (forwrite) {
2370 /* check if page is covered */
2371 sector_t sector = sh->dev[dd_idx].sector;
2372 for (bi=sh->dev[dd_idx].towrite;
2373 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2374 bi && bi->bi_sector <= sector;
2375 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2376 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2377 sector = bi->bi_sector + (bi->bi_size>>9);
2378 }
2379 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2380 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2381 }
2382 spin_unlock_irq(&conf->device_lock);
2383
2384 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2385 (unsigned long long)(*bip)->bi_sector,
2386 (unsigned long long)sh->sector, dd_idx);
2387
2388 if (conf->mddev->bitmap && firstwrite) {
2389 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2390 STRIPE_SECTORS, 0);
2391 sh->bm_seq = conf->seq_flush+1;
2392 set_bit(STRIPE_BIT_DELAY, &sh->state);
2393 }
2394 return 1;
2395
2396 overlap:
2397 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2398 spin_unlock_irq(&conf->device_lock);
2399 return 0;
2400}
2401
2402static void end_reshape(struct r5conf *conf);
2403
2404static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2405 struct stripe_head *sh)
2406{
2407 int sectors_per_chunk =
2408 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2409 int dd_idx;
2410 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2411 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2412
2413 raid5_compute_sector(conf,
2414 stripe * (disks - conf->max_degraded)
2415 *sectors_per_chunk + chunk_offset,
2416 previous,
2417 &dd_idx, sh);
2418}
2419
2420static void
2421handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2422 struct stripe_head_state *s, int disks,
2423 struct bio **return_bi)
2424{
2425 int i;
2426 for (i = disks; i--; ) {
2427 struct bio *bi;
2428 int bitmap_end = 0;
2429
2430 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2431 struct md_rdev *rdev;
2432 rcu_read_lock();
2433 rdev = rcu_dereference(conf->disks[i].rdev);
2434 if (rdev && test_bit(In_sync, &rdev->flags))
2435 atomic_inc(&rdev->nr_pending);
2436 else
2437 rdev = NULL;
2438 rcu_read_unlock();
2439 if (rdev) {
2440 if (!rdev_set_badblocks(
2441 rdev,
2442 sh->sector,
2443 STRIPE_SECTORS, 0))
2444 md_error(conf->mddev, rdev);
2445 rdev_dec_pending(rdev, conf->mddev);
2446 }
2447 }
2448 spin_lock_irq(&conf->device_lock);
2449 /* fail all writes first */
2450 bi = sh->dev[i].towrite;
2451 sh->dev[i].towrite = NULL;
2452 if (bi) {
2453 s->to_write--;
2454 bitmap_end = 1;
2455 }
2456
2457 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2458 wake_up(&conf->wait_for_overlap);
2459
2460 while (bi && bi->bi_sector <
2461 sh->dev[i].sector + STRIPE_SECTORS) {
2462 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2463 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2464 if (!raid5_dec_bi_phys_segments(bi)) {
2465 md_write_end(conf->mddev);
2466 bi->bi_next = *return_bi;
2467 *return_bi = bi;
2468 }
2469 bi = nextbi;
2470 }
2471 /* and fail all 'written' */
2472 bi = sh->dev[i].written;
2473 sh->dev[i].written = NULL;
2474 if (bi) bitmap_end = 1;
2475 while (bi && bi->bi_sector <
2476 sh->dev[i].sector + STRIPE_SECTORS) {
2477 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2478 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2479 if (!raid5_dec_bi_phys_segments(bi)) {
2480 md_write_end(conf->mddev);
2481 bi->bi_next = *return_bi;
2482 *return_bi = bi;
2483 }
2484 bi = bi2;
2485 }
2486
2487 /* fail any reads if this device is non-operational and
2488 * the data has not reached the cache yet.
2489 */
2490 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2491 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2492 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2493 bi = sh->dev[i].toread;
2494 sh->dev[i].toread = NULL;
2495 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2496 wake_up(&conf->wait_for_overlap);
2497 if (bi) s->to_read--;
2498 while (bi && bi->bi_sector <
2499 sh->dev[i].sector + STRIPE_SECTORS) {
2500 struct bio *nextbi =
2501 r5_next_bio(bi, sh->dev[i].sector);
2502 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2503 if (!raid5_dec_bi_phys_segments(bi)) {
2504 bi->bi_next = *return_bi;
2505 *return_bi = bi;
2506 }
2507 bi = nextbi;
2508 }
2509 }
2510 spin_unlock_irq(&conf->device_lock);
2511 if (bitmap_end)
2512 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2513 STRIPE_SECTORS, 0, 0);
2514 /* If we were in the middle of a write the parity block might
2515 * still be locked - so just clear all R5_LOCKED flags
2516 */
2517 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2518 }
2519
2520 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2521 if (atomic_dec_and_test(&conf->pending_full_writes))
2522 md_wakeup_thread(conf->mddev->thread);
2523}
2524
2525static void
2526handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2527 struct stripe_head_state *s)
2528{
2529 int abort = 0;
2530 int i;
2531
2532 clear_bit(STRIPE_SYNCING, &sh->state);
2533 s->syncing = 0;
2534 s->replacing = 0;
2535 /* There is nothing more to do for sync/check/repair.
2536 * Don't even need to abort as that is handled elsewhere
2537 * if needed, and not always wanted e.g. if there is a known
2538 * bad block here.
2539 * For recover/replace we need to record a bad block on all
2540 * non-sync devices, or abort the recovery
2541 */
2542 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2543 /* During recovery devices cannot be removed, so
2544 * locking and refcounting of rdevs is not needed
2545 */
2546 for (i = 0; i < conf->raid_disks; i++) {
2547 struct md_rdev *rdev = conf->disks[i].rdev;
2548 if (rdev
2549 && !test_bit(Faulty, &rdev->flags)
2550 && !test_bit(In_sync, &rdev->flags)
2551 && !rdev_set_badblocks(rdev, sh->sector,
2552 STRIPE_SECTORS, 0))
2553 abort = 1;
2554 rdev = conf->disks[i].replacement;
2555 if (rdev
2556 && !test_bit(Faulty, &rdev->flags)
2557 && !test_bit(In_sync, &rdev->flags)
2558 && !rdev_set_badblocks(rdev, sh->sector,
2559 STRIPE_SECTORS, 0))
2560 abort = 1;
2561 }
2562 if (abort)
2563 conf->recovery_disabled =
2564 conf->mddev->recovery_disabled;
2565 }
2566 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2567}
2568
2569static int want_replace(struct stripe_head *sh, int disk_idx)
2570{
2571 struct md_rdev *rdev;
2572 int rv = 0;
2573 /* Doing recovery so rcu locking not required */
2574 rdev = sh->raid_conf->disks[disk_idx].replacement;
2575 if (rdev
2576 && !test_bit(Faulty, &rdev->flags)
2577 && !test_bit(In_sync, &rdev->flags)
2578 && (rdev->recovery_offset <= sh->sector
2579 || rdev->mddev->recovery_cp <= sh->sector))
2580 rv = 1;
2581
2582 return rv;
2583}
2584
2585/* fetch_block - checks the given member device to see if its data needs
2586 * to be read or computed to satisfy a request.
2587 *
2588 * Returns 1 when no more member devices need to be checked, otherwise returns
2589 * 0 to tell the loop in handle_stripe_fill to continue
2590 */
2591static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2592 int disk_idx, int disks)
2593{
2594 struct r5dev *dev = &sh->dev[disk_idx];
2595 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2596 &sh->dev[s->failed_num[1]] };
2597
2598 /* is the data in this block needed, and can we get it? */
2599 if (!test_bit(R5_LOCKED, &dev->flags) &&
2600 !test_bit(R5_UPTODATE, &dev->flags) &&
2601 (dev->toread ||
2602 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2603 s->syncing || s->expanding ||
2604 (s->replacing && want_replace(sh, disk_idx)) ||
2605 (s->failed >= 1 && fdev[0]->toread) ||
2606 (s->failed >= 2 && fdev[1]->toread) ||
2607 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2608 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2609 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2610 /* we would like to get this block, possibly by computing it,
2611 * otherwise read it if the backing disk is insync
2612 */
2613 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2614 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2615 if ((s->uptodate == disks - 1) &&
2616 (s->failed && (disk_idx == s->failed_num[0] ||
2617 disk_idx == s->failed_num[1]))) {
2618 /* have disk failed, and we're requested to fetch it;
2619 * do compute it
2620 */
2621 pr_debug("Computing stripe %llu block %d\n",
2622 (unsigned long long)sh->sector, disk_idx);
2623 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2624 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2625 set_bit(R5_Wantcompute, &dev->flags);
2626 sh->ops.target = disk_idx;
2627 sh->ops.target2 = -1; /* no 2nd target */
2628 s->req_compute = 1;
2629 /* Careful: from this point on 'uptodate' is in the eye
2630 * of raid_run_ops which services 'compute' operations
2631 * before writes. R5_Wantcompute flags a block that will
2632 * be R5_UPTODATE by the time it is needed for a
2633 * subsequent operation.
2634 */
2635 s->uptodate++;
2636 return 1;
2637 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2638 /* Computing 2-failure is *very* expensive; only
2639 * do it if failed >= 2
2640 */
2641 int other;
2642 for (other = disks; other--; ) {
2643 if (other == disk_idx)
2644 continue;
2645 if (!test_bit(R5_UPTODATE,
2646 &sh->dev[other].flags))
2647 break;
2648 }
2649 BUG_ON(other < 0);
2650 pr_debug("Computing stripe %llu blocks %d,%d\n",
2651 (unsigned long long)sh->sector,
2652 disk_idx, other);
2653 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2654 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2655 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2656 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2657 sh->ops.target = disk_idx;
2658 sh->ops.target2 = other;
2659 s->uptodate += 2;
2660 s->req_compute = 1;
2661 return 1;
2662 } else if (test_bit(R5_Insync, &dev->flags)) {
2663 set_bit(R5_LOCKED, &dev->flags);
2664 set_bit(R5_Wantread, &dev->flags);
2665 s->locked++;
2666 pr_debug("Reading block %d (sync=%d)\n",
2667 disk_idx, s->syncing);
2668 }
2669 }
2670
2671 return 0;
2672}
2673
2674/**
2675 * handle_stripe_fill - read or compute data to satisfy pending requests.
2676 */
2677static void handle_stripe_fill(struct stripe_head *sh,
2678 struct stripe_head_state *s,
2679 int disks)
2680{
2681 int i;
2682
2683 /* look for blocks to read/compute, skip this if a compute
2684 * is already in flight, or if the stripe contents are in the
2685 * midst of changing due to a write
2686 */
2687 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2688 !sh->reconstruct_state)
2689 for (i = disks; i--; )
2690 if (fetch_block(sh, s, i, disks))
2691 break;
2692 set_bit(STRIPE_HANDLE, &sh->state);
2693}
2694
2695
2696/* handle_stripe_clean_event
2697 * any written block on an uptodate or failed drive can be returned.
2698 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2699 * never LOCKED, so we don't need to test 'failed' directly.
2700 */
2701static void handle_stripe_clean_event(struct r5conf *conf,
2702 struct stripe_head *sh, int disks, struct bio **return_bi)
2703{
2704 int i;
2705 struct r5dev *dev;
2706
2707 for (i = disks; i--; )
2708 if (sh->dev[i].written) {
2709 dev = &sh->dev[i];
2710 if (!test_bit(R5_LOCKED, &dev->flags) &&
2711 test_bit(R5_UPTODATE, &dev->flags)) {
2712 /* We can return any write requests */
2713 struct bio *wbi, *wbi2;
2714 int bitmap_end = 0;
2715 pr_debug("Return write for disc %d\n", i);
2716 spin_lock_irq(&conf->device_lock);
2717 wbi = dev->written;
2718 dev->written = NULL;
2719 while (wbi && wbi->bi_sector <
2720 dev->sector + STRIPE_SECTORS) {
2721 wbi2 = r5_next_bio(wbi, dev->sector);
2722 if (!raid5_dec_bi_phys_segments(wbi)) {
2723 md_write_end(conf->mddev);
2724 wbi->bi_next = *return_bi;
2725 *return_bi = wbi;
2726 }
2727 wbi = wbi2;
2728 }
2729 if (dev->towrite == NULL)
2730 bitmap_end = 1;
2731 spin_unlock_irq(&conf->device_lock);
2732 if (bitmap_end)
2733 bitmap_endwrite(conf->mddev->bitmap,
2734 sh->sector,
2735 STRIPE_SECTORS,
2736 !test_bit(STRIPE_DEGRADED, &sh->state),
2737 0);
2738 }
2739 }
2740
2741 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2742 if (atomic_dec_and_test(&conf->pending_full_writes))
2743 md_wakeup_thread(conf->mddev->thread);
2744}
2745
2746static void handle_stripe_dirtying(struct r5conf *conf,
2747 struct stripe_head *sh,
2748 struct stripe_head_state *s,
2749 int disks)
2750{
2751 int rmw = 0, rcw = 0, i;
2752 if (conf->max_degraded == 2) {
2753 /* RAID6 requires 'rcw' in current implementation
2754 * Calculate the real rcw later - for now fake it
2755 * look like rcw is cheaper
2756 */
2757 rcw = 1; rmw = 2;
2758 } else for (i = disks; i--; ) {
2759 /* would I have to read this buffer for read_modify_write */
2760 struct r5dev *dev = &sh->dev[i];
2761 if ((dev->towrite || i == sh->pd_idx) &&
2762 !test_bit(R5_LOCKED, &dev->flags) &&
2763 !(test_bit(R5_UPTODATE, &dev->flags) ||
2764 test_bit(R5_Wantcompute, &dev->flags))) {
2765 if (test_bit(R5_Insync, &dev->flags))
2766 rmw++;
2767 else
2768 rmw += 2*disks; /* cannot read it */
2769 }
2770 /* Would I have to read this buffer for reconstruct_write */
2771 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2772 !test_bit(R5_LOCKED, &dev->flags) &&
2773 !(test_bit(R5_UPTODATE, &dev->flags) ||
2774 test_bit(R5_Wantcompute, &dev->flags))) {
2775 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2776 else
2777 rcw += 2*disks;
2778 }
2779 }
2780 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2781 (unsigned long long)sh->sector, rmw, rcw);
2782 set_bit(STRIPE_HANDLE, &sh->state);
2783 if (rmw < rcw && rmw > 0)
2784 /* prefer read-modify-write, but need to get some data */
2785 for (i = disks; i--; ) {
2786 struct r5dev *dev = &sh->dev[i];
2787 if ((dev->towrite || i == sh->pd_idx) &&
2788 !test_bit(R5_LOCKED, &dev->flags) &&
2789 !(test_bit(R5_UPTODATE, &dev->flags) ||
2790 test_bit(R5_Wantcompute, &dev->flags)) &&
2791 test_bit(R5_Insync, &dev->flags)) {
2792 if (
2793 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2794 pr_debug("Read_old block "
2795 "%d for r-m-w\n", i);
2796 set_bit(R5_LOCKED, &dev->flags);
2797 set_bit(R5_Wantread, &dev->flags);
2798 s->locked++;
2799 } else {
2800 set_bit(STRIPE_DELAYED, &sh->state);
2801 set_bit(STRIPE_HANDLE, &sh->state);
2802 }
2803 }
2804 }
2805 if (rcw <= rmw && rcw > 0) {
2806 /* want reconstruct write, but need to get some data */
2807 rcw = 0;
2808 for (i = disks; i--; ) {
2809 struct r5dev *dev = &sh->dev[i];
2810 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2811 i != sh->pd_idx && i != sh->qd_idx &&
2812 !test_bit(R5_LOCKED, &dev->flags) &&
2813 !(test_bit(R5_UPTODATE, &dev->flags) ||
2814 test_bit(R5_Wantcompute, &dev->flags))) {
2815 rcw++;
2816 if (!test_bit(R5_Insync, &dev->flags))
2817 continue; /* it's a failed drive */
2818 if (
2819 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2820 pr_debug("Read_old block "
2821 "%d for Reconstruct\n", i);
2822 set_bit(R5_LOCKED, &dev->flags);
2823 set_bit(R5_Wantread, &dev->flags);
2824 s->locked++;
2825 } else {
2826 set_bit(STRIPE_DELAYED, &sh->state);
2827 set_bit(STRIPE_HANDLE, &sh->state);
2828 }
2829 }
2830 }
2831 }
2832 /* now if nothing is locked, and if we have enough data,
2833 * we can start a write request
2834 */
2835 /* since handle_stripe can be called at any time we need to handle the
2836 * case where a compute block operation has been submitted and then a
2837 * subsequent call wants to start a write request. raid_run_ops only
2838 * handles the case where compute block and reconstruct are requested
2839 * simultaneously. If this is not the case then new writes need to be
2840 * held off until the compute completes.
2841 */
2842 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2843 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2844 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2845 schedule_reconstruction(sh, s, rcw == 0, 0);
2846}
2847
2848static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2849 struct stripe_head_state *s, int disks)
2850{
2851 struct r5dev *dev = NULL;
2852
2853 set_bit(STRIPE_HANDLE, &sh->state);
2854
2855 switch (sh->check_state) {
2856 case check_state_idle:
2857 /* start a new check operation if there are no failures */
2858 if (s->failed == 0) {
2859 BUG_ON(s->uptodate != disks);
2860 sh->check_state = check_state_run;
2861 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2862 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2863 s->uptodate--;
2864 break;
2865 }
2866 dev = &sh->dev[s->failed_num[0]];
2867 /* fall through */
2868 case check_state_compute_result:
2869 sh->check_state = check_state_idle;
2870 if (!dev)
2871 dev = &sh->dev[sh->pd_idx];
2872
2873 /* check that a write has not made the stripe insync */
2874 if (test_bit(STRIPE_INSYNC, &sh->state))
2875 break;
2876
2877 /* either failed parity check, or recovery is happening */
2878 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2879 BUG_ON(s->uptodate != disks);
2880
2881 set_bit(R5_LOCKED, &dev->flags);
2882 s->locked++;
2883 set_bit(R5_Wantwrite, &dev->flags);
2884
2885 clear_bit(STRIPE_DEGRADED, &sh->state);
2886 set_bit(STRIPE_INSYNC, &sh->state);
2887 break;
2888 case check_state_run:
2889 break; /* we will be called again upon completion */
2890 case check_state_check_result:
2891 sh->check_state = check_state_idle;
2892
2893 /* if a failure occurred during the check operation, leave
2894 * STRIPE_INSYNC not set and let the stripe be handled again
2895 */
2896 if (s->failed)
2897 break;
2898
2899 /* handle a successful check operation, if parity is correct
2900 * we are done. Otherwise update the mismatch count and repair
2901 * parity if !MD_RECOVERY_CHECK
2902 */
2903 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2904 /* parity is correct (on disc,
2905 * not in buffer any more)
2906 */
2907 set_bit(STRIPE_INSYNC, &sh->state);
2908 else {
2909 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2910 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2911 /* don't try to repair!! */
2912 set_bit(STRIPE_INSYNC, &sh->state);
2913 else {
2914 sh->check_state = check_state_compute_run;
2915 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2916 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2917 set_bit(R5_Wantcompute,
2918 &sh->dev[sh->pd_idx].flags);
2919 sh->ops.target = sh->pd_idx;
2920 sh->ops.target2 = -1;
2921 s->uptodate++;
2922 }
2923 }
2924 break;
2925 case check_state_compute_run:
2926 break;
2927 default:
2928 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2929 __func__, sh->check_state,
2930 (unsigned long long) sh->sector);
2931 BUG();
2932 }
2933}
2934
2935
2936static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2937 struct stripe_head_state *s,
2938 int disks)
2939{
2940 int pd_idx = sh->pd_idx;
2941 int qd_idx = sh->qd_idx;
2942 struct r5dev *dev;
2943
2944 set_bit(STRIPE_HANDLE, &sh->state);
2945
2946 BUG_ON(s->failed > 2);
2947
2948 /* Want to check and possibly repair P and Q.
2949 * However there could be one 'failed' device, in which
2950 * case we can only check one of them, possibly using the
2951 * other to generate missing data
2952 */
2953
2954 switch (sh->check_state) {
2955 case check_state_idle:
2956 /* start a new check operation if there are < 2 failures */
2957 if (s->failed == s->q_failed) {
2958 /* The only possible failed device holds Q, so it
2959 * makes sense to check P (If anything else were failed,
2960 * we would have used P to recreate it).
2961 */
2962 sh->check_state = check_state_run;
2963 }
2964 if (!s->q_failed && s->failed < 2) {
2965 /* Q is not failed, and we didn't use it to generate
2966 * anything, so it makes sense to check it
2967 */
2968 if (sh->check_state == check_state_run)
2969 sh->check_state = check_state_run_pq;
2970 else
2971 sh->check_state = check_state_run_q;
2972 }
2973
2974 /* discard potentially stale zero_sum_result */
2975 sh->ops.zero_sum_result = 0;
2976
2977 if (sh->check_state == check_state_run) {
2978 /* async_xor_zero_sum destroys the contents of P */
2979 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2980 s->uptodate--;
2981 }
2982 if (sh->check_state >= check_state_run &&
2983 sh->check_state <= check_state_run_pq) {
2984 /* async_syndrome_zero_sum preserves P and Q, so
2985 * no need to mark them !uptodate here
2986 */
2987 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2988 break;
2989 }
2990
2991 /* we have 2-disk failure */
2992 BUG_ON(s->failed != 2);
2993 /* fall through */
2994 case check_state_compute_result:
2995 sh->check_state = check_state_idle;
2996
2997 /* check that a write has not made the stripe insync */
2998 if (test_bit(STRIPE_INSYNC, &sh->state))
2999 break;
3000
3001 /* now write out any block on a failed drive,
3002 * or P or Q if they were recomputed
3003 */
3004 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3005 if (s->failed == 2) {
3006 dev = &sh->dev[s->failed_num[1]];
3007 s->locked++;
3008 set_bit(R5_LOCKED, &dev->flags);
3009 set_bit(R5_Wantwrite, &dev->flags);
3010 }
3011 if (s->failed >= 1) {
3012 dev = &sh->dev[s->failed_num[0]];
3013 s->locked++;
3014 set_bit(R5_LOCKED, &dev->flags);
3015 set_bit(R5_Wantwrite, &dev->flags);
3016 }
3017 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3018 dev = &sh->dev[pd_idx];
3019 s->locked++;
3020 set_bit(R5_LOCKED, &dev->flags);
3021 set_bit(R5_Wantwrite, &dev->flags);
3022 }
3023 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3024 dev = &sh->dev[qd_idx];
3025 s->locked++;
3026 set_bit(R5_LOCKED, &dev->flags);
3027 set_bit(R5_Wantwrite, &dev->flags);
3028 }
3029 clear_bit(STRIPE_DEGRADED, &sh->state);
3030
3031 set_bit(STRIPE_INSYNC, &sh->state);
3032 break;
3033 case check_state_run:
3034 case check_state_run_q:
3035 case check_state_run_pq:
3036 break; /* we will be called again upon completion */
3037 case check_state_check_result:
3038 sh->check_state = check_state_idle;
3039
3040 /* handle a successful check operation, if parity is correct
3041 * we are done. Otherwise update the mismatch count and repair
3042 * parity if !MD_RECOVERY_CHECK
3043 */
3044 if (sh->ops.zero_sum_result == 0) {
3045 /* both parities are correct */
3046 if (!s->failed)
3047 set_bit(STRIPE_INSYNC, &sh->state);
3048 else {
3049 /* in contrast to the raid5 case we can validate
3050 * parity, but still have a failure to write
3051 * back
3052 */
3053 sh->check_state = check_state_compute_result;
3054 /* Returning at this point means that we may go
3055 * off and bring p and/or q uptodate again so
3056 * we make sure to check zero_sum_result again
3057 * to verify if p or q need writeback
3058 */
3059 }
3060 } else {
3061 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3062 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3063 /* don't try to repair!! */
3064 set_bit(STRIPE_INSYNC, &sh->state);
3065 else {
3066 int *target = &sh->ops.target;
3067
3068 sh->ops.target = -1;
3069 sh->ops.target2 = -1;
3070 sh->check_state = check_state_compute_run;
3071 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3072 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3073 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3074 set_bit(R5_Wantcompute,
3075 &sh->dev[pd_idx].flags);
3076 *target = pd_idx;
3077 target = &sh->ops.target2;
3078 s->uptodate++;
3079 }
3080 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3081 set_bit(R5_Wantcompute,
3082 &sh->dev[qd_idx].flags);
3083 *target = qd_idx;
3084 s->uptodate++;
3085 }
3086 }
3087 }
3088 break;
3089 case check_state_compute_run:
3090 break;
3091 default:
3092 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3093 __func__, sh->check_state,
3094 (unsigned long long) sh->sector);
3095 BUG();
3096 }
3097}
3098
3099static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3100{
3101 int i;
3102
3103 /* We have read all the blocks in this stripe and now we need to
3104 * copy some of them into a target stripe for expand.
3105 */
3106 struct dma_async_tx_descriptor *tx = NULL;
3107 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3108 for (i = 0; i < sh->disks; i++)
3109 if (i != sh->pd_idx && i != sh->qd_idx) {
3110 int dd_idx, j;
3111 struct stripe_head *sh2;
3112 struct async_submit_ctl submit;
3113
3114 sector_t bn = compute_blocknr(sh, i, 1);
3115 sector_t s = raid5_compute_sector(conf, bn, 0,
3116 &dd_idx, NULL);
3117 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3118 if (sh2 == NULL)
3119 /* so far only the early blocks of this stripe
3120 * have been requested. When later blocks
3121 * get requested, we will try again
3122 */
3123 continue;
3124 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3125 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3126 /* must have already done this block */
3127 release_stripe(sh2);
3128 continue;
3129 }
3130
3131 /* place all the copies on one channel */
3132 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3133 tx = async_memcpy(sh2->dev[dd_idx].page,
3134 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3135 &submit);
3136
3137 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3138 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3139 for (j = 0; j < conf->raid_disks; j++)
3140 if (j != sh2->pd_idx &&
3141 j != sh2->qd_idx &&
3142 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3143 break;
3144 if (j == conf->raid_disks) {
3145 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3146 set_bit(STRIPE_HANDLE, &sh2->state);
3147 }
3148 release_stripe(sh2);
3149
3150 }
3151 /* done submitting copies, wait for them to complete */
3152 if (tx) {
3153 async_tx_ack(tx);
3154 dma_wait_for_async_tx(tx);
3155 }
3156}
3157
3158/*
3159 * handle_stripe - do things to a stripe.
3160 *
3161 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3162 * state of various bits to see what needs to be done.
3163 * Possible results:
3164 * return some read requests which now have data
3165 * return some write requests which are safely on storage
3166 * schedule a read on some buffers
3167 * schedule a write of some buffers
3168 * return confirmation of parity correctness
3169 *
3170 */
3171
3172static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3173{
3174 struct r5conf *conf = sh->raid_conf;
3175 int disks = sh->disks;
3176 struct r5dev *dev;
3177 int i;
3178 int do_recovery = 0;
3179
3180 memset(s, 0, sizeof(*s));
3181
3182 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3183 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3184 s->failed_num[0] = -1;
3185 s->failed_num[1] = -1;
3186
3187 /* Now to look around and see what can be done */
3188 rcu_read_lock();
3189 spin_lock_irq(&conf->device_lock);
3190 for (i=disks; i--; ) {
3191 struct md_rdev *rdev;
3192 sector_t first_bad;
3193 int bad_sectors;
3194 int is_bad = 0;
3195
3196 dev = &sh->dev[i];
3197
3198 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3199 i, dev->flags,
3200 dev->toread, dev->towrite, dev->written);
3201 /* maybe we can reply to a read
3202 *
3203 * new wantfill requests are only permitted while
3204 * ops_complete_biofill is guaranteed to be inactive
3205 */
3206 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3207 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3208 set_bit(R5_Wantfill, &dev->flags);
3209
3210 /* now count some things */
3211 if (test_bit(R5_LOCKED, &dev->flags))
3212 s->locked++;
3213 if (test_bit(R5_UPTODATE, &dev->flags))
3214 s->uptodate++;
3215 if (test_bit(R5_Wantcompute, &dev->flags)) {
3216 s->compute++;
3217 BUG_ON(s->compute > 2);
3218 }
3219
3220 if (test_bit(R5_Wantfill, &dev->flags))
3221 s->to_fill++;
3222 else if (dev->toread)
3223 s->to_read++;
3224 if (dev->towrite) {
3225 s->to_write++;
3226 if (!test_bit(R5_OVERWRITE, &dev->flags))
3227 s->non_overwrite++;
3228 }
3229 if (dev->written)
3230 s->written++;
3231 /* Prefer to use the replacement for reads, but only
3232 * if it is recovered enough and has no bad blocks.
3233 */
3234 rdev = rcu_dereference(conf->disks[i].replacement);
3235 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3236 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3237 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3238 &first_bad, &bad_sectors))
3239 set_bit(R5_ReadRepl, &dev->flags);
3240 else {
3241 if (rdev)
3242 set_bit(R5_NeedReplace, &dev->flags);
3243 rdev = rcu_dereference(conf->disks[i].rdev);
3244 clear_bit(R5_ReadRepl, &dev->flags);
3245 }
3246 if (rdev && test_bit(Faulty, &rdev->flags))
3247 rdev = NULL;
3248 if (rdev) {
3249 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3250 &first_bad, &bad_sectors);
3251 if (s->blocked_rdev == NULL
3252 && (test_bit(Blocked, &rdev->flags)
3253 || is_bad < 0)) {
3254 if (is_bad < 0)
3255 set_bit(BlockedBadBlocks,
3256 &rdev->flags);
3257 s->blocked_rdev = rdev;
3258 atomic_inc(&rdev->nr_pending);
3259 }
3260 }
3261 clear_bit(R5_Insync, &dev->flags);
3262 if (!rdev)
3263 /* Not in-sync */;
3264 else if (is_bad) {
3265 /* also not in-sync */
3266 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3267 test_bit(R5_UPTODATE, &dev->flags)) {
3268 /* treat as in-sync, but with a read error
3269 * which we can now try to correct
3270 */
3271 set_bit(R5_Insync, &dev->flags);
3272 set_bit(R5_ReadError, &dev->flags);
3273 }
3274 } else if (test_bit(In_sync, &rdev->flags))
3275 set_bit(R5_Insync, &dev->flags);
3276 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3277 /* in sync if before recovery_offset */
3278 set_bit(R5_Insync, &dev->flags);
3279 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3280 test_bit(R5_Expanded, &dev->flags))
3281 /* If we've reshaped into here, we assume it is Insync.
3282 * We will shortly update recovery_offset to make
3283 * it official.
3284 */
3285 set_bit(R5_Insync, &dev->flags);
3286
3287 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3288 /* This flag does not apply to '.replacement'
3289 * only to .rdev, so make sure to check that*/
3290 struct md_rdev *rdev2 = rcu_dereference(
3291 conf->disks[i].rdev);
3292 if (rdev2 == rdev)
3293 clear_bit(R5_Insync, &dev->flags);
3294 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3295 s->handle_bad_blocks = 1;
3296 atomic_inc(&rdev2->nr_pending);
3297 } else
3298 clear_bit(R5_WriteError, &dev->flags);
3299 }
3300 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3301 /* This flag does not apply to '.replacement'
3302 * only to .rdev, so make sure to check that*/
3303 struct md_rdev *rdev2 = rcu_dereference(
3304 conf->disks[i].rdev);
3305 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3306 s->handle_bad_blocks = 1;
3307 atomic_inc(&rdev2->nr_pending);
3308 } else
3309 clear_bit(R5_MadeGood, &dev->flags);
3310 }
3311 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3312 struct md_rdev *rdev2 = rcu_dereference(
3313 conf->disks[i].replacement);
3314 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3315 s->handle_bad_blocks = 1;
3316 atomic_inc(&rdev2->nr_pending);
3317 } else
3318 clear_bit(R5_MadeGoodRepl, &dev->flags);
3319 }
3320 if (!test_bit(R5_Insync, &dev->flags)) {
3321 /* The ReadError flag will just be confusing now */
3322 clear_bit(R5_ReadError, &dev->flags);
3323 clear_bit(R5_ReWrite, &dev->flags);
3324 }
3325 if (test_bit(R5_ReadError, &dev->flags))
3326 clear_bit(R5_Insync, &dev->flags);
3327 if (!test_bit(R5_Insync, &dev->flags)) {
3328 if (s->failed < 2)
3329 s->failed_num[s->failed] = i;
3330 s->failed++;
3331 if (rdev && !test_bit(Faulty, &rdev->flags))
3332 do_recovery = 1;
3333 }
3334 }
3335 spin_unlock_irq(&conf->device_lock);
3336 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3337 /* If there is a failed device being replaced,
3338 * we must be recovering.
3339 * else if we are after recovery_cp, we must be syncing
3340 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3341 * else we can only be replacing
3342 * sync and recovery both need to read all devices, and so
3343 * use the same flag.
3344 */
3345 if (do_recovery ||
3346 sh->sector >= conf->mddev->recovery_cp ||
3347 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3348 s->syncing = 1;
3349 else
3350 s->replacing = 1;
3351 }
3352 rcu_read_unlock();
3353}
3354
3355static void handle_stripe(struct stripe_head *sh)
3356{
3357 struct stripe_head_state s;
3358 struct r5conf *conf = sh->raid_conf;
3359 int i;
3360 int prexor;
3361 int disks = sh->disks;
3362 struct r5dev *pdev, *qdev;
3363
3364 clear_bit(STRIPE_HANDLE, &sh->state);
3365 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3366 /* already being handled, ensure it gets handled
3367 * again when current action finishes */
3368 set_bit(STRIPE_HANDLE, &sh->state);
3369 return;
3370 }
3371
3372 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3373 set_bit(STRIPE_SYNCING, &sh->state);
3374 clear_bit(STRIPE_INSYNC, &sh->state);
3375 }
3376 clear_bit(STRIPE_DELAYED, &sh->state);
3377
3378 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3379 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3380 (unsigned long long)sh->sector, sh->state,
3381 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3382 sh->check_state, sh->reconstruct_state);
3383
3384 analyse_stripe(sh, &s);
3385
3386 if (s.handle_bad_blocks) {
3387 set_bit(STRIPE_HANDLE, &sh->state);
3388 goto finish;
3389 }
3390
3391 if (unlikely(s.blocked_rdev)) {
3392 if (s.syncing || s.expanding || s.expanded ||
3393 s.replacing || s.to_write || s.written) {
3394 set_bit(STRIPE_HANDLE, &sh->state);
3395 goto finish;
3396 }
3397 /* There is nothing for the blocked_rdev to block */
3398 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3399 s.blocked_rdev = NULL;
3400 }
3401
3402 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3403 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3404 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3405 }
3406
3407 pr_debug("locked=%d uptodate=%d to_read=%d"
3408 " to_write=%d failed=%d failed_num=%d,%d\n",
3409 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3410 s.failed_num[0], s.failed_num[1]);
3411 /* check if the array has lost more than max_degraded devices and,
3412 * if so, some requests might need to be failed.
3413 */
3414 if (s.failed > conf->max_degraded) {
3415 sh->check_state = 0;
3416 sh->reconstruct_state = 0;
3417 if (s.to_read+s.to_write+s.written)
3418 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3419 if (s.syncing + s.replacing)
3420 handle_failed_sync(conf, sh, &s);
3421 }
3422
3423 /*
3424 * might be able to return some write requests if the parity blocks
3425 * are safe, or on a failed drive
3426 */
3427 pdev = &sh->dev[sh->pd_idx];
3428 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3429 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3430 qdev = &sh->dev[sh->qd_idx];
3431 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3432 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3433 || conf->level < 6;
3434
3435 if (s.written &&
3436 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3437 && !test_bit(R5_LOCKED, &pdev->flags)
3438 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3439 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3440 && !test_bit(R5_LOCKED, &qdev->flags)
3441 && test_bit(R5_UPTODATE, &qdev->flags)))))
3442 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3443
3444 /* Now we might consider reading some blocks, either to check/generate
3445 * parity, or to satisfy requests
3446 * or to load a block that is being partially written.
3447 */
3448 if (s.to_read || s.non_overwrite
3449 || (conf->level == 6 && s.to_write && s.failed)
3450 || (s.syncing && (s.uptodate + s.compute < disks))
3451 || s.replacing
3452 || s.expanding)
3453 handle_stripe_fill(sh, &s, disks);
3454
3455 /* Now we check to see if any write operations have recently
3456 * completed
3457 */
3458 prexor = 0;
3459 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3460 prexor = 1;
3461 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3462 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3463 sh->reconstruct_state = reconstruct_state_idle;
3464
3465 /* All the 'written' buffers and the parity block are ready to
3466 * be written back to disk
3467 */
3468 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3469 BUG_ON(sh->qd_idx >= 0 &&
3470 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3471 for (i = disks; i--; ) {
3472 struct r5dev *dev = &sh->dev[i];
3473 if (test_bit(R5_LOCKED, &dev->flags) &&
3474 (i == sh->pd_idx || i == sh->qd_idx ||
3475 dev->written)) {
3476 pr_debug("Writing block %d\n", i);
3477 set_bit(R5_Wantwrite, &dev->flags);
3478 if (prexor)
3479 continue;
3480 if (!test_bit(R5_Insync, &dev->flags) ||
3481 ((i == sh->pd_idx || i == sh->qd_idx) &&
3482 s.failed == 0))
3483 set_bit(STRIPE_INSYNC, &sh->state);
3484 }
3485 }
3486 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3487 s.dec_preread_active = 1;
3488 }
3489
3490 /* Now to consider new write requests and what else, if anything
3491 * should be read. We do not handle new writes when:
3492 * 1/ A 'write' operation (copy+xor) is already in flight.
3493 * 2/ A 'check' operation is in flight, as it may clobber the parity
3494 * block.
3495 */
3496 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3497 handle_stripe_dirtying(conf, sh, &s, disks);
3498
3499 /* maybe we need to check and possibly fix the parity for this stripe
3500 * Any reads will already have been scheduled, so we just see if enough
3501 * data is available. The parity check is held off while parity
3502 * dependent operations are in flight.
3503 */
3504 if (sh->check_state ||
3505 (s.syncing && s.locked == 0 &&
3506 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3507 !test_bit(STRIPE_INSYNC, &sh->state))) {
3508 if (conf->level == 6)
3509 handle_parity_checks6(conf, sh, &s, disks);
3510 else
3511 handle_parity_checks5(conf, sh, &s, disks);
3512 }
3513
3514 if (s.replacing && s.locked == 0
3515 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3516 /* Write out to replacement devices where possible */
3517 for (i = 0; i < conf->raid_disks; i++)
3518 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3519 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3520 set_bit(R5_WantReplace, &sh->dev[i].flags);
3521 set_bit(R5_LOCKED, &sh->dev[i].flags);
3522 s.locked++;
3523 }
3524 set_bit(STRIPE_INSYNC, &sh->state);
3525 }
3526 if ((s.syncing || s.replacing) && s.locked == 0 &&
3527 test_bit(STRIPE_INSYNC, &sh->state)) {
3528 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3529 clear_bit(STRIPE_SYNCING, &sh->state);
3530 }
3531
3532 /* If the failed drives are just a ReadError, then we might need
3533 * to progress the repair/check process
3534 */
3535 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3536 for (i = 0; i < s.failed; i++) {
3537 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3538 if (test_bit(R5_ReadError, &dev->flags)
3539 && !test_bit(R5_LOCKED, &dev->flags)
3540 && test_bit(R5_UPTODATE, &dev->flags)
3541 ) {
3542 if (!test_bit(R5_ReWrite, &dev->flags)) {
3543 set_bit(R5_Wantwrite, &dev->flags);
3544 set_bit(R5_ReWrite, &dev->flags);
3545 set_bit(R5_LOCKED, &dev->flags);
3546 s.locked++;
3547 } else {
3548 /* let's read it back */
3549 set_bit(R5_Wantread, &dev->flags);
3550 set_bit(R5_LOCKED, &dev->flags);
3551 s.locked++;
3552 }
3553 }
3554 }
3555
3556
3557 /* Finish reconstruct operations initiated by the expansion process */
3558 if (sh->reconstruct_state == reconstruct_state_result) {
3559 struct stripe_head *sh_src
3560 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3561 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3562 /* sh cannot be written until sh_src has been read.
3563 * so arrange for sh to be delayed a little
3564 */
3565 set_bit(STRIPE_DELAYED, &sh->state);
3566 set_bit(STRIPE_HANDLE, &sh->state);
3567 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3568 &sh_src->state))
3569 atomic_inc(&conf->preread_active_stripes);
3570 release_stripe(sh_src);
3571 goto finish;
3572 }
3573 if (sh_src)
3574 release_stripe(sh_src);
3575
3576 sh->reconstruct_state = reconstruct_state_idle;
3577 clear_bit(STRIPE_EXPANDING, &sh->state);
3578 for (i = conf->raid_disks; i--; ) {
3579 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3580 set_bit(R5_LOCKED, &sh->dev[i].flags);
3581 s.locked++;
3582 }
3583 }
3584
3585 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3586 !sh->reconstruct_state) {
3587 /* Need to write out all blocks after computing parity */
3588 sh->disks = conf->raid_disks;
3589 stripe_set_idx(sh->sector, conf, 0, sh);
3590 schedule_reconstruction(sh, &s, 1, 1);
3591 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3592 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3593 atomic_dec(&conf->reshape_stripes);
3594 wake_up(&conf->wait_for_overlap);
3595 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3596 }
3597
3598 if (s.expanding && s.locked == 0 &&
3599 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3600 handle_stripe_expansion(conf, sh);
3601
3602finish:
3603 /* wait for this device to become unblocked */
3604 if (unlikely(s.blocked_rdev)) {
3605 if (conf->mddev->external)
3606 md_wait_for_blocked_rdev(s.blocked_rdev,
3607 conf->mddev);
3608 else
3609 /* Internal metadata will immediately
3610 * be written by raid5d, so we don't
3611 * need to wait here.
3612 */
3613 rdev_dec_pending(s.blocked_rdev,
3614 conf->mddev);
3615 }
3616
3617 if (s.handle_bad_blocks)
3618 for (i = disks; i--; ) {
3619 struct md_rdev *rdev;
3620 struct r5dev *dev = &sh->dev[i];
3621 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3622 /* We own a safe reference to the rdev */
3623 rdev = conf->disks[i].rdev;
3624 if (!rdev_set_badblocks(rdev, sh->sector,
3625 STRIPE_SECTORS, 0))
3626 md_error(conf->mddev, rdev);
3627 rdev_dec_pending(rdev, conf->mddev);
3628 }
3629 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3630 rdev = conf->disks[i].rdev;
3631 rdev_clear_badblocks(rdev, sh->sector,
3632 STRIPE_SECTORS, 0);
3633 rdev_dec_pending(rdev, conf->mddev);
3634 }
3635 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3636 rdev = conf->disks[i].replacement;
3637 if (!rdev)
3638 /* rdev have been moved down */
3639 rdev = conf->disks[i].rdev;
3640 rdev_clear_badblocks(rdev, sh->sector,
3641 STRIPE_SECTORS, 0);
3642 rdev_dec_pending(rdev, conf->mddev);
3643 }
3644 }
3645
3646 if (s.ops_request)
3647 raid_run_ops(sh, s.ops_request);
3648
3649 ops_run_io(sh, &s);
3650
3651 if (s.dec_preread_active) {
3652 /* We delay this until after ops_run_io so that if make_request
3653 * is waiting on a flush, it won't continue until the writes
3654 * have actually been submitted.
3655 */
3656 atomic_dec(&conf->preread_active_stripes);
3657 if (atomic_read(&conf->preread_active_stripes) <
3658 IO_THRESHOLD)
3659 md_wakeup_thread(conf->mddev->thread);
3660 }
3661
3662 return_io(s.return_bi);
3663
3664 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3665}
3666
3667static void raid5_activate_delayed(struct r5conf *conf)
3668{
3669 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3670 while (!list_empty(&conf->delayed_list)) {
3671 struct list_head *l = conf->delayed_list.next;
3672 struct stripe_head *sh;
3673 sh = list_entry(l, struct stripe_head, lru);
3674 list_del_init(l);
3675 clear_bit(STRIPE_DELAYED, &sh->state);
3676 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3677 atomic_inc(&conf->preread_active_stripes);
3678 list_add_tail(&sh->lru, &conf->hold_list);
3679 }
3680 }
3681}
3682
3683static void activate_bit_delay(struct r5conf *conf)
3684{
3685 /* device_lock is held */
3686 struct list_head head;
3687 list_add(&head, &conf->bitmap_list);
3688 list_del_init(&conf->bitmap_list);
3689 while (!list_empty(&head)) {
3690 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3691 list_del_init(&sh->lru);
3692 atomic_inc(&sh->count);
3693 __release_stripe(conf, sh);
3694 }
3695}
3696
3697int md_raid5_congested(struct mddev *mddev, int bits)
3698{
3699 struct r5conf *conf = mddev->private;
3700
3701 /* No difference between reads and writes. Just check
3702 * how busy the stripe_cache is
3703 */
3704
3705 if (conf->inactive_blocked)
3706 return 1;
3707 if (conf->quiesce)
3708 return 1;
3709 if (list_empty_careful(&conf->inactive_list))
3710 return 1;
3711
3712 return 0;
3713}
3714EXPORT_SYMBOL_GPL(md_raid5_congested);
3715
3716static int raid5_congested(void *data, int bits)
3717{
3718 struct mddev *mddev = data;
3719
3720 return mddev_congested(mddev, bits) ||
3721 md_raid5_congested(mddev, bits);
3722}
3723
3724/* We want read requests to align with chunks where possible,
3725 * but write requests don't need to.
3726 */
3727static int raid5_mergeable_bvec(struct request_queue *q,
3728 struct bvec_merge_data *bvm,
3729 struct bio_vec *biovec)
3730{
3731 struct mddev *mddev = q->queuedata;
3732 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3733 int max;
3734 unsigned int chunk_sectors = mddev->chunk_sectors;
3735 unsigned int bio_sectors = bvm->bi_size >> 9;
3736
3737 if ((bvm->bi_rw & 1) == WRITE)
3738 return biovec->bv_len; /* always allow writes to be mergeable */
3739
3740 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3741 chunk_sectors = mddev->new_chunk_sectors;
3742 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3743 if (max < 0) max = 0;
3744 if (max <= biovec->bv_len && bio_sectors == 0)
3745 return biovec->bv_len;
3746 else
3747 return max;
3748}
3749
3750
3751static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3752{
3753 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3754 unsigned int chunk_sectors = mddev->chunk_sectors;
3755 unsigned int bio_sectors = bio->bi_size >> 9;
3756
3757 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3758 chunk_sectors = mddev->new_chunk_sectors;
3759 return chunk_sectors >=
3760 ((sector & (chunk_sectors - 1)) + bio_sectors);
3761}
3762
3763/*
3764 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3765 * later sampled by raid5d.
3766 */
3767static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3768{
3769 unsigned long flags;
3770
3771 spin_lock_irqsave(&conf->device_lock, flags);
3772
3773 bi->bi_next = conf->retry_read_aligned_list;
3774 conf->retry_read_aligned_list = bi;
3775
3776 spin_unlock_irqrestore(&conf->device_lock, flags);
3777 md_wakeup_thread(conf->mddev->thread);
3778}
3779
3780
3781static struct bio *remove_bio_from_retry(struct r5conf *conf)
3782{
3783 struct bio *bi;
3784
3785 bi = conf->retry_read_aligned;
3786 if (bi) {
3787 conf->retry_read_aligned = NULL;
3788 return bi;
3789 }
3790 bi = conf->retry_read_aligned_list;
3791 if(bi) {
3792 conf->retry_read_aligned_list = bi->bi_next;
3793 bi->bi_next = NULL;
3794 /*
3795 * this sets the active strip count to 1 and the processed
3796 * strip count to zero (upper 8 bits)
3797 */
3798 bi->bi_phys_segments = 1; /* biased count of active stripes */
3799 }
3800
3801 return bi;
3802}
3803
3804
3805/*
3806 * The "raid5_align_endio" should check if the read succeeded and if it
3807 * did, call bio_endio on the original bio (having bio_put the new bio
3808 * first).
3809 * If the read failed..
3810 */
3811static void raid5_align_endio(struct bio *bi, int error)
3812{
3813 struct bio* raid_bi = bi->bi_private;
3814 struct mddev *mddev;
3815 struct r5conf *conf;
3816 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3817 struct md_rdev *rdev;
3818
3819 bio_put(bi);
3820
3821 rdev = (void*)raid_bi->bi_next;
3822 raid_bi->bi_next = NULL;
3823 mddev = rdev->mddev;
3824 conf = mddev->private;
3825
3826 rdev_dec_pending(rdev, conf->mddev);
3827
3828 if (!error && uptodate) {
3829 bio_endio(raid_bi, 0);
3830 if (atomic_dec_and_test(&conf->active_aligned_reads))
3831 wake_up(&conf->wait_for_stripe);
3832 return;
3833 }
3834
3835
3836 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3837
3838 add_bio_to_retry(raid_bi, conf);
3839}
3840
3841static int bio_fits_rdev(struct bio *bi)
3842{
3843 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3844
3845 if ((bi->bi_size>>9) > queue_max_sectors(q))
3846 return 0;
3847 blk_recount_segments(q, bi);
3848 if (bi->bi_phys_segments > queue_max_segments(q))
3849 return 0;
3850
3851 if (q->merge_bvec_fn)
3852 /* it's too hard to apply the merge_bvec_fn at this stage,
3853 * just just give up
3854 */
3855 return 0;
3856
3857 return 1;
3858}
3859
3860
3861static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3862{
3863 struct r5conf *conf = mddev->private;
3864 int dd_idx;
3865 struct bio* align_bi;
3866 struct md_rdev *rdev;
3867 sector_t end_sector;
3868
3869 if (!in_chunk_boundary(mddev, raid_bio)) {
3870 pr_debug("chunk_aligned_read : non aligned\n");
3871 return 0;
3872 }
3873 /*
3874 * use bio_clone_mddev to make a copy of the bio
3875 */
3876 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3877 if (!align_bi)
3878 return 0;
3879 /*
3880 * set bi_end_io to a new function, and set bi_private to the
3881 * original bio.
3882 */
3883 align_bi->bi_end_io = raid5_align_endio;
3884 align_bi->bi_private = raid_bio;
3885 /*
3886 * compute position
3887 */
3888 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3889 0,
3890 &dd_idx, NULL);
3891
3892 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3893 rcu_read_lock();
3894 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3895 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3896 rdev->recovery_offset < end_sector) {
3897 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3898 if (rdev &&
3899 (test_bit(Faulty, &rdev->flags) ||
3900 !(test_bit(In_sync, &rdev->flags) ||
3901 rdev->recovery_offset >= end_sector)))
3902 rdev = NULL;
3903 }
3904 if (rdev) {
3905 sector_t first_bad;
3906 int bad_sectors;
3907
3908 atomic_inc(&rdev->nr_pending);
3909 rcu_read_unlock();
3910 raid_bio->bi_next = (void*)rdev;
3911 align_bi->bi_bdev = rdev->bdev;
3912 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3913
3914 if (!bio_fits_rdev(align_bi) ||
3915 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3916 &first_bad, &bad_sectors)) {
3917 /* too big in some way, or has a known bad block */
3918 bio_put(align_bi);
3919 rdev_dec_pending(rdev, mddev);
3920 return 0;
3921 }
3922
3923 /* No reshape active, so we can trust rdev->data_offset */
3924 align_bi->bi_sector += rdev->data_offset;
3925
3926 spin_lock_irq(&conf->device_lock);
3927 wait_event_lock_irq(conf->wait_for_stripe,
3928 conf->quiesce == 0,
3929 conf->device_lock, /* nothing */);
3930 atomic_inc(&conf->active_aligned_reads);
3931 spin_unlock_irq(&conf->device_lock);
3932
3933 generic_make_request(align_bi);
3934 return 1;
3935 } else {
3936 rcu_read_unlock();
3937 bio_put(align_bi);
3938 return 0;
3939 }
3940}
3941
3942/* __get_priority_stripe - get the next stripe to process
3943 *
3944 * Full stripe writes are allowed to pass preread active stripes up until
3945 * the bypass_threshold is exceeded. In general the bypass_count
3946 * increments when the handle_list is handled before the hold_list; however, it
3947 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3948 * stripe with in flight i/o. The bypass_count will be reset when the
3949 * head of the hold_list has changed, i.e. the head was promoted to the
3950 * handle_list.
3951 */
3952static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3953{
3954 struct stripe_head *sh;
3955
3956 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3957 __func__,
3958 list_empty(&conf->handle_list) ? "empty" : "busy",
3959 list_empty(&conf->hold_list) ? "empty" : "busy",
3960 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3961
3962 if (!list_empty(&conf->handle_list)) {
3963 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3964
3965 if (list_empty(&conf->hold_list))
3966 conf->bypass_count = 0;
3967 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3968 if (conf->hold_list.next == conf->last_hold)
3969 conf->bypass_count++;
3970 else {
3971 conf->last_hold = conf->hold_list.next;
3972 conf->bypass_count -= conf->bypass_threshold;
3973 if (conf->bypass_count < 0)
3974 conf->bypass_count = 0;
3975 }
3976 }
3977 } else if (!list_empty(&conf->hold_list) &&
3978 ((conf->bypass_threshold &&
3979 conf->bypass_count > conf->bypass_threshold) ||
3980 atomic_read(&conf->pending_full_writes) == 0)) {
3981 sh = list_entry(conf->hold_list.next,
3982 typeof(*sh), lru);
3983 conf->bypass_count -= conf->bypass_threshold;
3984 if (conf->bypass_count < 0)
3985 conf->bypass_count = 0;
3986 } else
3987 return NULL;
3988
3989 list_del_init(&sh->lru);
3990 atomic_inc(&sh->count);
3991 BUG_ON(atomic_read(&sh->count) != 1);
3992 return sh;
3993}
3994
3995static void make_request(struct mddev *mddev, struct bio * bi)
3996{
3997 struct r5conf *conf = mddev->private;
3998 int dd_idx;
3999 sector_t new_sector;
4000 sector_t logical_sector, last_sector;
4001 struct stripe_head *sh;
4002 const int rw = bio_data_dir(bi);
4003 int remaining;
4004
4005 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4006 md_flush_request(mddev, bi);
4007 return;
4008 }
4009
4010 md_write_start(mddev, bi);
4011
4012 if (rw == READ &&
4013 mddev->reshape_position == MaxSector &&
4014 chunk_aligned_read(mddev,bi))
4015 return;
4016
4017 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4018 last_sector = bi->bi_sector + (bi->bi_size>>9);
4019 bi->bi_next = NULL;
4020 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4021
4022 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4023 DEFINE_WAIT(w);
4024 int previous;
4025
4026 retry:
4027 previous = 0;
4028 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4029 if (unlikely(conf->reshape_progress != MaxSector)) {
4030 /* spinlock is needed as reshape_progress may be
4031 * 64bit on a 32bit platform, and so it might be
4032 * possible to see a half-updated value
4033 * Of course reshape_progress could change after
4034 * the lock is dropped, so once we get a reference
4035 * to the stripe that we think it is, we will have
4036 * to check again.
4037 */
4038 spin_lock_irq(&conf->device_lock);
4039 if (mddev->reshape_backwards
4040 ? logical_sector < conf->reshape_progress
4041 : logical_sector >= conf->reshape_progress) {
4042 previous = 1;
4043 } else {
4044 if (mddev->reshape_backwards
4045 ? logical_sector < conf->reshape_safe
4046 : logical_sector >= conf->reshape_safe) {
4047 spin_unlock_irq(&conf->device_lock);
4048 schedule();
4049 goto retry;
4050 }
4051 }
4052 spin_unlock_irq(&conf->device_lock);
4053 }
4054
4055 new_sector = raid5_compute_sector(conf, logical_sector,
4056 previous,
4057 &dd_idx, NULL);
4058 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4059 (unsigned long long)new_sector,
4060 (unsigned long long)logical_sector);
4061
4062 sh = get_active_stripe(conf, new_sector, previous,
4063 (bi->bi_rw&RWA_MASK), 0);
4064 if (sh) {
4065 if (unlikely(previous)) {
4066 /* expansion might have moved on while waiting for a
4067 * stripe, so we must do the range check again.
4068 * Expansion could still move past after this
4069 * test, but as we are holding a reference to
4070 * 'sh', we know that if that happens,
4071 * STRIPE_EXPANDING will get set and the expansion
4072 * won't proceed until we finish with the stripe.
4073 */
4074 int must_retry = 0;
4075 spin_lock_irq(&conf->device_lock);
4076 if (mddev->reshape_backwards
4077 ? logical_sector >= conf->reshape_progress
4078 : logical_sector < conf->reshape_progress)
4079 /* mismatch, need to try again */
4080 must_retry = 1;
4081 spin_unlock_irq(&conf->device_lock);
4082 if (must_retry) {
4083 release_stripe(sh);
4084 schedule();
4085 goto retry;
4086 }
4087 }
4088
4089 if (rw == WRITE &&
4090 logical_sector >= mddev->suspend_lo &&
4091 logical_sector < mddev->suspend_hi) {
4092 release_stripe(sh);
4093 /* As the suspend_* range is controlled by
4094 * userspace, we want an interruptible
4095 * wait.
4096 */
4097 flush_signals(current);
4098 prepare_to_wait(&conf->wait_for_overlap,
4099 &w, TASK_INTERRUPTIBLE);
4100 if (logical_sector >= mddev->suspend_lo &&
4101 logical_sector < mddev->suspend_hi)
4102 schedule();
4103 goto retry;
4104 }
4105
4106 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4107 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4108 /* Stripe is busy expanding or
4109 * add failed due to overlap. Flush everything
4110 * and wait a while
4111 */
4112 md_wakeup_thread(mddev->thread);
4113 release_stripe(sh);
4114 schedule();
4115 goto retry;
4116 }
4117 finish_wait(&conf->wait_for_overlap, &w);
4118 set_bit(STRIPE_HANDLE, &sh->state);
4119 clear_bit(STRIPE_DELAYED, &sh->state);
4120 if ((bi->bi_rw & REQ_SYNC) &&
4121 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4122 atomic_inc(&conf->preread_active_stripes);
4123 mddev_check_plugged(mddev);
4124 release_stripe(sh);
4125 } else {
4126 /* cannot get stripe for read-ahead, just give-up */
4127 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4128 finish_wait(&conf->wait_for_overlap, &w);
4129 break;
4130 }
4131 }
4132
4133 spin_lock_irq(&conf->device_lock);
4134 remaining = raid5_dec_bi_phys_segments(bi);
4135 spin_unlock_irq(&conf->device_lock);
4136 if (remaining == 0) {
4137
4138 if ( rw == WRITE )
4139 md_write_end(mddev);
4140
4141 bio_endio(bi, 0);
4142 }
4143}
4144
4145static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4146
4147static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4148{
4149 /* reshaping is quite different to recovery/resync so it is
4150 * handled quite separately ... here.
4151 *
4152 * On each call to sync_request, we gather one chunk worth of
4153 * destination stripes and flag them as expanding.
4154 * Then we find all the source stripes and request reads.
4155 * As the reads complete, handle_stripe will copy the data
4156 * into the destination stripe and release that stripe.
4157 */
4158 struct r5conf *conf = mddev->private;
4159 struct stripe_head *sh;
4160 sector_t first_sector, last_sector;
4161 int raid_disks = conf->previous_raid_disks;
4162 int data_disks = raid_disks - conf->max_degraded;
4163 int new_data_disks = conf->raid_disks - conf->max_degraded;
4164 int i;
4165 int dd_idx;
4166 sector_t writepos, readpos, safepos;
4167 sector_t stripe_addr;
4168 int reshape_sectors;
4169 struct list_head stripes;
4170
4171 if (sector_nr == 0) {
4172 /* If restarting in the middle, skip the initial sectors */
4173 if (mddev->reshape_backwards &&
4174 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4175 sector_nr = raid5_size(mddev, 0, 0)
4176 - conf->reshape_progress;
4177 } else if (!mddev->reshape_backwards &&
4178 conf->reshape_progress > 0)
4179 sector_nr = conf->reshape_progress;
4180 sector_div(sector_nr, new_data_disks);
4181 if (sector_nr) {
4182 mddev->curr_resync_completed = sector_nr;
4183 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4184 *skipped = 1;
4185 return sector_nr;
4186 }
4187 }
4188
4189 /* We need to process a full chunk at a time.
4190 * If old and new chunk sizes differ, we need to process the
4191 * largest of these
4192 */
4193 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4194 reshape_sectors = mddev->new_chunk_sectors;
4195 else
4196 reshape_sectors = mddev->chunk_sectors;
4197
4198 /* We update the metadata at least every 10 seconds, or when
4199 * the data about to be copied would over-write the source of
4200 * the data at the front of the range. i.e. one new_stripe
4201 * along from reshape_progress new_maps to after where
4202 * reshape_safe old_maps to
4203 */
4204 writepos = conf->reshape_progress;
4205 sector_div(writepos, new_data_disks);
4206 readpos = conf->reshape_progress;
4207 sector_div(readpos, data_disks);
4208 safepos = conf->reshape_safe;
4209 sector_div(safepos, data_disks);
4210 if (mddev->reshape_backwards) {
4211 writepos -= min_t(sector_t, reshape_sectors, writepos);
4212 readpos += reshape_sectors;
4213 safepos += reshape_sectors;
4214 } else {
4215 writepos += reshape_sectors;
4216 readpos -= min_t(sector_t, reshape_sectors, readpos);
4217 safepos -= min_t(sector_t, reshape_sectors, safepos);
4218 }
4219
4220 /* Having calculated the 'writepos' possibly use it
4221 * to set 'stripe_addr' which is where we will write to.
4222 */
4223 if (mddev->reshape_backwards) {
4224 BUG_ON(conf->reshape_progress == 0);
4225 stripe_addr = writepos;
4226 BUG_ON((mddev->dev_sectors &
4227 ~((sector_t)reshape_sectors - 1))
4228 - reshape_sectors - stripe_addr
4229 != sector_nr);
4230 } else {
4231 BUG_ON(writepos != sector_nr + reshape_sectors);
4232 stripe_addr = sector_nr;
4233 }
4234
4235 /* 'writepos' is the most advanced device address we might write.
4236 * 'readpos' is the least advanced device address we might read.
4237 * 'safepos' is the least address recorded in the metadata as having
4238 * been reshaped.
4239 * If there is a min_offset_diff, these are adjusted either by
4240 * increasing the safepos/readpos if diff is negative, or
4241 * increasing writepos if diff is positive.
4242 * If 'readpos' is then behind 'writepos', there is no way that we can
4243 * ensure safety in the face of a crash - that must be done by userspace
4244 * making a backup of the data. So in that case there is no particular
4245 * rush to update metadata.
4246 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4247 * update the metadata to advance 'safepos' to match 'readpos' so that
4248 * we can be safe in the event of a crash.
4249 * So we insist on updating metadata if safepos is behind writepos and
4250 * readpos is beyond writepos.
4251 * In any case, update the metadata every 10 seconds.
4252 * Maybe that number should be configurable, but I'm not sure it is
4253 * worth it.... maybe it could be a multiple of safemode_delay???
4254 */
4255 if (conf->min_offset_diff < 0) {
4256 safepos += -conf->min_offset_diff;
4257 readpos += -conf->min_offset_diff;
4258 } else
4259 writepos += conf->min_offset_diff;
4260
4261 if ((mddev->reshape_backwards
4262 ? (safepos > writepos && readpos < writepos)
4263 : (safepos < writepos && readpos > writepos)) ||
4264 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4265 /* Cannot proceed until we've updated the superblock... */
4266 wait_event(conf->wait_for_overlap,
4267 atomic_read(&conf->reshape_stripes)==0);
4268 mddev->reshape_position = conf->reshape_progress;
4269 mddev->curr_resync_completed = sector_nr;
4270 conf->reshape_checkpoint = jiffies;
4271 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4272 md_wakeup_thread(mddev->thread);
4273 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4274 kthread_should_stop());
4275 spin_lock_irq(&conf->device_lock);
4276 conf->reshape_safe = mddev->reshape_position;
4277 spin_unlock_irq(&conf->device_lock);
4278 wake_up(&conf->wait_for_overlap);
4279 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4280 }
4281
4282 INIT_LIST_HEAD(&stripes);
4283 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4284 int j;
4285 int skipped_disk = 0;
4286 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4287 set_bit(STRIPE_EXPANDING, &sh->state);
4288 atomic_inc(&conf->reshape_stripes);
4289 /* If any of this stripe is beyond the end of the old
4290 * array, then we need to zero those blocks
4291 */
4292 for (j=sh->disks; j--;) {
4293 sector_t s;
4294 if (j == sh->pd_idx)
4295 continue;
4296 if (conf->level == 6 &&
4297 j == sh->qd_idx)
4298 continue;
4299 s = compute_blocknr(sh, j, 0);
4300 if (s < raid5_size(mddev, 0, 0)) {
4301 skipped_disk = 1;
4302 continue;
4303 }
4304 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4305 set_bit(R5_Expanded, &sh->dev[j].flags);
4306 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4307 }
4308 if (!skipped_disk) {
4309 set_bit(STRIPE_EXPAND_READY, &sh->state);
4310 set_bit(STRIPE_HANDLE, &sh->state);
4311 }
4312 list_add(&sh->lru, &stripes);
4313 }
4314 spin_lock_irq(&conf->device_lock);
4315 if (mddev->reshape_backwards)
4316 conf->reshape_progress -= reshape_sectors * new_data_disks;
4317 else
4318 conf->reshape_progress += reshape_sectors * new_data_disks;
4319 spin_unlock_irq(&conf->device_lock);
4320 /* Ok, those stripe are ready. We can start scheduling
4321 * reads on the source stripes.
4322 * The source stripes are determined by mapping the first and last
4323 * block on the destination stripes.
4324 */
4325 first_sector =
4326 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4327 1, &dd_idx, NULL);
4328 last_sector =
4329 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4330 * new_data_disks - 1),
4331 1, &dd_idx, NULL);
4332 if (last_sector >= mddev->dev_sectors)
4333 last_sector = mddev->dev_sectors - 1;
4334 while (first_sector <= last_sector) {
4335 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4336 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4337 set_bit(STRIPE_HANDLE, &sh->state);
4338 release_stripe(sh);
4339 first_sector += STRIPE_SECTORS;
4340 }
4341 /* Now that the sources are clearly marked, we can release
4342 * the destination stripes
4343 */
4344 while (!list_empty(&stripes)) {
4345 sh = list_entry(stripes.next, struct stripe_head, lru);
4346 list_del_init(&sh->lru);
4347 release_stripe(sh);
4348 }
4349 /* If this takes us to the resync_max point where we have to pause,
4350 * then we need to write out the superblock.
4351 */
4352 sector_nr += reshape_sectors;
4353 if ((sector_nr - mddev->curr_resync_completed) * 2
4354 >= mddev->resync_max - mddev->curr_resync_completed) {
4355 /* Cannot proceed until we've updated the superblock... */
4356 wait_event(conf->wait_for_overlap,
4357 atomic_read(&conf->reshape_stripes) == 0);
4358 mddev->reshape_position = conf->reshape_progress;
4359 mddev->curr_resync_completed = sector_nr;
4360 conf->reshape_checkpoint = jiffies;
4361 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4362 md_wakeup_thread(mddev->thread);
4363 wait_event(mddev->sb_wait,
4364 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4365 || kthread_should_stop());
4366 spin_lock_irq(&conf->device_lock);
4367 conf->reshape_safe = mddev->reshape_position;
4368 spin_unlock_irq(&conf->device_lock);
4369 wake_up(&conf->wait_for_overlap);
4370 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4371 }
4372 return reshape_sectors;
4373}
4374
4375/* FIXME go_faster isn't used */
4376static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4377{
4378 struct r5conf *conf = mddev->private;
4379 struct stripe_head *sh;
4380 sector_t max_sector = mddev->dev_sectors;
4381 sector_t sync_blocks;
4382 int still_degraded = 0;
4383 int i;
4384
4385 if (sector_nr >= max_sector) {
4386 /* just being told to finish up .. nothing much to do */
4387
4388 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4389 end_reshape(conf);
4390 return 0;
4391 }
4392
4393 if (mddev->curr_resync < max_sector) /* aborted */
4394 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4395 &sync_blocks, 1);
4396 else /* completed sync */
4397 conf->fullsync = 0;
4398 bitmap_close_sync(mddev->bitmap);
4399
4400 return 0;
4401 }
4402
4403 /* Allow raid5_quiesce to complete */
4404 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4405
4406 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4407 return reshape_request(mddev, sector_nr, skipped);
4408
4409 /* No need to check resync_max as we never do more than one
4410 * stripe, and as resync_max will always be on a chunk boundary,
4411 * if the check in md_do_sync didn't fire, there is no chance
4412 * of overstepping resync_max here
4413 */
4414
4415 /* if there is too many failed drives and we are trying
4416 * to resync, then assert that we are finished, because there is
4417 * nothing we can do.
4418 */
4419 if (mddev->degraded >= conf->max_degraded &&
4420 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4421 sector_t rv = mddev->dev_sectors - sector_nr;
4422 *skipped = 1;
4423 return rv;
4424 }
4425 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4426 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4427 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4428 /* we can skip this block, and probably more */
4429 sync_blocks /= STRIPE_SECTORS;
4430 *skipped = 1;
4431 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4432 }
4433
4434 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4435
4436 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4437 if (sh == NULL) {
4438 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4439 /* make sure we don't swamp the stripe cache if someone else
4440 * is trying to get access
4441 */
4442 schedule_timeout_uninterruptible(1);
4443 }
4444 /* Need to check if array will still be degraded after recovery/resync
4445 * We don't need to check the 'failed' flag as when that gets set,
4446 * recovery aborts.
4447 */
4448 for (i = 0; i < conf->raid_disks; i++)
4449 if (conf->disks[i].rdev == NULL)
4450 still_degraded = 1;
4451
4452 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4453
4454 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4455
4456 handle_stripe(sh);
4457 release_stripe(sh);
4458
4459 return STRIPE_SECTORS;
4460}
4461
4462static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4463{
4464 /* We may not be able to submit a whole bio at once as there
4465 * may not be enough stripe_heads available.
4466 * We cannot pre-allocate enough stripe_heads as we may need
4467 * more than exist in the cache (if we allow ever large chunks).
4468 * So we do one stripe head at a time and record in
4469 * ->bi_hw_segments how many have been done.
4470 *
4471 * We *know* that this entire raid_bio is in one chunk, so
4472 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4473 */
4474 struct stripe_head *sh;
4475 int dd_idx;
4476 sector_t sector, logical_sector, last_sector;
4477 int scnt = 0;
4478 int remaining;
4479 int handled = 0;
4480
4481 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4482 sector = raid5_compute_sector(conf, logical_sector,
4483 0, &dd_idx, NULL);
4484 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4485
4486 for (; logical_sector < last_sector;
4487 logical_sector += STRIPE_SECTORS,
4488 sector += STRIPE_SECTORS,
4489 scnt++) {
4490
4491 if (scnt < raid5_bi_hw_segments(raid_bio))
4492 /* already done this stripe */
4493 continue;
4494
4495 sh = get_active_stripe(conf, sector, 0, 1, 0);
4496
4497 if (!sh) {
4498 /* failed to get a stripe - must wait */
4499 raid5_set_bi_hw_segments(raid_bio, scnt);
4500 conf->retry_read_aligned = raid_bio;
4501 return handled;
4502 }
4503
4504 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4505 release_stripe(sh);
4506 raid5_set_bi_hw_segments(raid_bio, scnt);
4507 conf->retry_read_aligned = raid_bio;
4508 return handled;
4509 }
4510
4511 handle_stripe(sh);
4512 release_stripe(sh);
4513 handled++;
4514 }
4515 spin_lock_irq(&conf->device_lock);
4516 remaining = raid5_dec_bi_phys_segments(raid_bio);
4517 spin_unlock_irq(&conf->device_lock);
4518 if (remaining == 0)
4519 bio_endio(raid_bio, 0);
4520 if (atomic_dec_and_test(&conf->active_aligned_reads))
4521 wake_up(&conf->wait_for_stripe);
4522 return handled;
4523}
4524
4525
4526/*
4527 * This is our raid5 kernel thread.
4528 *
4529 * We scan the hash table for stripes which can be handled now.
4530 * During the scan, completed stripes are saved for us by the interrupt
4531 * handler, so that they will not have to wait for our next wakeup.
4532 */
4533static void raid5d(struct mddev *mddev)
4534{
4535 struct stripe_head *sh;
4536 struct r5conf *conf = mddev->private;
4537 int handled;
4538 struct blk_plug plug;
4539
4540 pr_debug("+++ raid5d active\n");
4541
4542 md_check_recovery(mddev);
4543
4544 blk_start_plug(&plug);
4545 handled = 0;
4546 spin_lock_irq(&conf->device_lock);
4547 while (1) {
4548 struct bio *bio;
4549
4550 if (atomic_read(&mddev->plug_cnt) == 0 &&
4551 !list_empty(&conf->bitmap_list)) {
4552 /* Now is a good time to flush some bitmap updates */
4553 conf->seq_flush++;
4554 spin_unlock_irq(&conf->device_lock);
4555 bitmap_unplug(mddev->bitmap);
4556 spin_lock_irq(&conf->device_lock);
4557 conf->seq_write = conf->seq_flush;
4558 activate_bit_delay(conf);
4559 }
4560 if (atomic_read(&mddev->plug_cnt) == 0)
4561 raid5_activate_delayed(conf);
4562
4563 while ((bio = remove_bio_from_retry(conf))) {
4564 int ok;
4565 spin_unlock_irq(&conf->device_lock);
4566 ok = retry_aligned_read(conf, bio);
4567 spin_lock_irq(&conf->device_lock);
4568 if (!ok)
4569 break;
4570 handled++;
4571 }
4572
4573 sh = __get_priority_stripe(conf);
4574
4575 if (!sh)
4576 break;
4577 spin_unlock_irq(&conf->device_lock);
4578
4579 handled++;
4580 handle_stripe(sh);
4581 release_stripe(sh);
4582 cond_resched();
4583
4584 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4585 md_check_recovery(mddev);
4586
4587 spin_lock_irq(&conf->device_lock);
4588 }
4589 pr_debug("%d stripes handled\n", handled);
4590
4591 spin_unlock_irq(&conf->device_lock);
4592
4593 async_tx_issue_pending_all();
4594 blk_finish_plug(&plug);
4595
4596 pr_debug("--- raid5d inactive\n");
4597}
4598
4599static ssize_t
4600raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4601{
4602 struct r5conf *conf = mddev->private;
4603 if (conf)
4604 return sprintf(page, "%d\n", conf->max_nr_stripes);
4605 else
4606 return 0;
4607}
4608
4609int
4610raid5_set_cache_size(struct mddev *mddev, int size)
4611{
4612 struct r5conf *conf = mddev->private;
4613 int err;
4614
4615 if (size <= 16 || size > 32768)
4616 return -EINVAL;
4617 while (size < conf->max_nr_stripes) {
4618 if (drop_one_stripe(conf))
4619 conf->max_nr_stripes--;
4620 else
4621 break;
4622 }
4623 err = md_allow_write(mddev);
4624 if (err)
4625 return err;
4626 while (size > conf->max_nr_stripes) {
4627 if (grow_one_stripe(conf))
4628 conf->max_nr_stripes++;
4629 else break;
4630 }
4631 return 0;
4632}
4633EXPORT_SYMBOL(raid5_set_cache_size);
4634
4635static ssize_t
4636raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4637{
4638 struct r5conf *conf = mddev->private;
4639 unsigned long new;
4640 int err;
4641
4642 if (len >= PAGE_SIZE)
4643 return -EINVAL;
4644 if (!conf)
4645 return -ENODEV;
4646
4647 if (strict_strtoul(page, 10, &new))
4648 return -EINVAL;
4649 err = raid5_set_cache_size(mddev, new);
4650 if (err)
4651 return err;
4652 return len;
4653}
4654
4655static struct md_sysfs_entry
4656raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4657 raid5_show_stripe_cache_size,
4658 raid5_store_stripe_cache_size);
4659
4660static ssize_t
4661raid5_show_preread_threshold(struct mddev *mddev, char *page)
4662{
4663 struct r5conf *conf = mddev->private;
4664 if (conf)
4665 return sprintf(page, "%d\n", conf->bypass_threshold);
4666 else
4667 return 0;
4668}
4669
4670static ssize_t
4671raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4672{
4673 struct r5conf *conf = mddev->private;
4674 unsigned long new;
4675 if (len >= PAGE_SIZE)
4676 return -EINVAL;
4677 if (!conf)
4678 return -ENODEV;
4679
4680 if (strict_strtoul(page, 10, &new))
4681 return -EINVAL;
4682 if (new > conf->max_nr_stripes)
4683 return -EINVAL;
4684 conf->bypass_threshold = new;
4685 return len;
4686}
4687
4688static struct md_sysfs_entry
4689raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4690 S_IRUGO | S_IWUSR,
4691 raid5_show_preread_threshold,
4692 raid5_store_preread_threshold);
4693
4694static ssize_t
4695stripe_cache_active_show(struct mddev *mddev, char *page)
4696{
4697 struct r5conf *conf = mddev->private;
4698 if (conf)
4699 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4700 else
4701 return 0;
4702}
4703
4704static struct md_sysfs_entry
4705raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4706
4707static struct attribute *raid5_attrs[] = {
4708 &raid5_stripecache_size.attr,
4709 &raid5_stripecache_active.attr,
4710 &raid5_preread_bypass_threshold.attr,
4711 NULL,
4712};
4713static struct attribute_group raid5_attrs_group = {
4714 .name = NULL,
4715 .attrs = raid5_attrs,
4716};
4717
4718static sector_t
4719raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4720{
4721 struct r5conf *conf = mddev->private;
4722
4723 if (!sectors)
4724 sectors = mddev->dev_sectors;
4725 if (!raid_disks)
4726 /* size is defined by the smallest of previous and new size */
4727 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4728
4729 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4730 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4731 return sectors * (raid_disks - conf->max_degraded);
4732}
4733
4734static void raid5_free_percpu(struct r5conf *conf)
4735{
4736 struct raid5_percpu *percpu;
4737 unsigned long cpu;
4738
4739 if (!conf->percpu)
4740 return;
4741
4742 get_online_cpus();
4743 for_each_possible_cpu(cpu) {
4744 percpu = per_cpu_ptr(conf->percpu, cpu);
4745 safe_put_page(percpu->spare_page);
4746 kfree(percpu->scribble);
4747 }
4748#ifdef CONFIG_HOTPLUG_CPU
4749 unregister_cpu_notifier(&conf->cpu_notify);
4750#endif
4751 put_online_cpus();
4752
4753 free_percpu(conf->percpu);
4754}
4755
4756static void free_conf(struct r5conf *conf)
4757{
4758 shrink_stripes(conf);
4759 raid5_free_percpu(conf);
4760 kfree(conf->disks);
4761 kfree(conf->stripe_hashtbl);
4762 kfree(conf);
4763}
4764
4765#ifdef CONFIG_HOTPLUG_CPU
4766static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4767 void *hcpu)
4768{
4769 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4770 long cpu = (long)hcpu;
4771 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4772
4773 switch (action) {
4774 case CPU_UP_PREPARE:
4775 case CPU_UP_PREPARE_FROZEN:
4776 if (conf->level == 6 && !percpu->spare_page)
4777 percpu->spare_page = alloc_page(GFP_KERNEL);
4778 if (!percpu->scribble)
4779 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4780
4781 if (!percpu->scribble ||
4782 (conf->level == 6 && !percpu->spare_page)) {
4783 safe_put_page(percpu->spare_page);
4784 kfree(percpu->scribble);
4785 pr_err("%s: failed memory allocation for cpu%ld\n",
4786 __func__, cpu);
4787 return notifier_from_errno(-ENOMEM);
4788 }
4789 break;
4790 case CPU_DEAD:
4791 case CPU_DEAD_FROZEN:
4792 safe_put_page(percpu->spare_page);
4793 kfree(percpu->scribble);
4794 percpu->spare_page = NULL;
4795 percpu->scribble = NULL;
4796 break;
4797 default:
4798 break;
4799 }
4800 return NOTIFY_OK;
4801}
4802#endif
4803
4804static int raid5_alloc_percpu(struct r5conf *conf)
4805{
4806 unsigned long cpu;
4807 struct page *spare_page;
4808 struct raid5_percpu __percpu *allcpus;
4809 void *scribble;
4810 int err;
4811
4812 allcpus = alloc_percpu(struct raid5_percpu);
4813 if (!allcpus)
4814 return -ENOMEM;
4815 conf->percpu = allcpus;
4816
4817 get_online_cpus();
4818 err = 0;
4819 for_each_present_cpu(cpu) {
4820 if (conf->level == 6) {
4821 spare_page = alloc_page(GFP_KERNEL);
4822 if (!spare_page) {
4823 err = -ENOMEM;
4824 break;
4825 }
4826 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4827 }
4828 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4829 if (!scribble) {
4830 err = -ENOMEM;
4831 break;
4832 }
4833 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4834 }
4835#ifdef CONFIG_HOTPLUG_CPU
4836 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4837 conf->cpu_notify.priority = 0;
4838 if (err == 0)
4839 err = register_cpu_notifier(&conf->cpu_notify);
4840#endif
4841 put_online_cpus();
4842
4843 return err;
4844}
4845
4846static struct r5conf *setup_conf(struct mddev *mddev)
4847{
4848 struct r5conf *conf;
4849 int raid_disk, memory, max_disks;
4850 struct md_rdev *rdev;
4851 struct disk_info *disk;
4852 char pers_name[6];
4853
4854 if (mddev->new_level != 5
4855 && mddev->new_level != 4
4856 && mddev->new_level != 6) {
4857 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4858 mdname(mddev), mddev->new_level);
4859 return ERR_PTR(-EIO);
4860 }
4861 if ((mddev->new_level == 5
4862 && !algorithm_valid_raid5(mddev->new_layout)) ||
4863 (mddev->new_level == 6
4864 && !algorithm_valid_raid6(mddev->new_layout))) {
4865 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4866 mdname(mddev), mddev->new_layout);
4867 return ERR_PTR(-EIO);
4868 }
4869 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4870 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4871 mdname(mddev), mddev->raid_disks);
4872 return ERR_PTR(-EINVAL);
4873 }
4874
4875 if (!mddev->new_chunk_sectors ||
4876 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4877 !is_power_of_2(mddev->new_chunk_sectors)) {
4878 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4879 mdname(mddev), mddev->new_chunk_sectors << 9);
4880 return ERR_PTR(-EINVAL);
4881 }
4882
4883 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4884 if (conf == NULL)
4885 goto abort;
4886 spin_lock_init(&conf->device_lock);
4887 init_waitqueue_head(&conf->wait_for_stripe);
4888 init_waitqueue_head(&conf->wait_for_overlap);
4889 INIT_LIST_HEAD(&conf->handle_list);
4890 INIT_LIST_HEAD(&conf->hold_list);
4891 INIT_LIST_HEAD(&conf->delayed_list);
4892 INIT_LIST_HEAD(&conf->bitmap_list);
4893 INIT_LIST_HEAD(&conf->inactive_list);
4894 atomic_set(&conf->active_stripes, 0);
4895 atomic_set(&conf->preread_active_stripes, 0);
4896 atomic_set(&conf->active_aligned_reads, 0);
4897 conf->bypass_threshold = BYPASS_THRESHOLD;
4898 conf->recovery_disabled = mddev->recovery_disabled - 1;
4899
4900 conf->raid_disks = mddev->raid_disks;
4901 if (mddev->reshape_position == MaxSector)
4902 conf->previous_raid_disks = mddev->raid_disks;
4903 else
4904 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4905 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4906 conf->scribble_len = scribble_len(max_disks);
4907
4908 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4909 GFP_KERNEL);
4910 if (!conf->disks)
4911 goto abort;
4912
4913 conf->mddev = mddev;
4914
4915 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4916 goto abort;
4917
4918 conf->level = mddev->new_level;
4919 if (raid5_alloc_percpu(conf) != 0)
4920 goto abort;
4921
4922 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4923
4924 rdev_for_each(rdev, mddev) {
4925 raid_disk = rdev->raid_disk;
4926 if (raid_disk >= max_disks
4927 || raid_disk < 0)
4928 continue;
4929 disk = conf->disks + raid_disk;
4930
4931 if (test_bit(Replacement, &rdev->flags)) {
4932 if (disk->replacement)
4933 goto abort;
4934 disk->replacement = rdev;
4935 } else {
4936 if (disk->rdev)
4937 goto abort;
4938 disk->rdev = rdev;
4939 }
4940
4941 if (test_bit(In_sync, &rdev->flags)) {
4942 char b[BDEVNAME_SIZE];
4943 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4944 " disk %d\n",
4945 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4946 } else if (rdev->saved_raid_disk != raid_disk)
4947 /* Cannot rely on bitmap to complete recovery */
4948 conf->fullsync = 1;
4949 }
4950
4951 conf->chunk_sectors = mddev->new_chunk_sectors;
4952 conf->level = mddev->new_level;
4953 if (conf->level == 6)
4954 conf->max_degraded = 2;
4955 else
4956 conf->max_degraded = 1;
4957 conf->algorithm = mddev->new_layout;
4958 conf->max_nr_stripes = NR_STRIPES;
4959 conf->reshape_progress = mddev->reshape_position;
4960 if (conf->reshape_progress != MaxSector) {
4961 conf->prev_chunk_sectors = mddev->chunk_sectors;
4962 conf->prev_algo = mddev->layout;
4963 }
4964
4965 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4966 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4967 if (grow_stripes(conf, conf->max_nr_stripes)) {
4968 printk(KERN_ERR
4969 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4970 mdname(mddev), memory);
4971 goto abort;
4972 } else
4973 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4974 mdname(mddev), memory);
4975
4976 sprintf(pers_name, "raid%d", mddev->new_level);
4977 conf->thread = md_register_thread(raid5d, mddev, pers_name);
4978 if (!conf->thread) {
4979 printk(KERN_ERR
4980 "md/raid:%s: couldn't allocate thread.\n",
4981 mdname(mddev));
4982 goto abort;
4983 }
4984
4985 return conf;
4986
4987 abort:
4988 if (conf) {
4989 free_conf(conf);
4990 return ERR_PTR(-EIO);
4991 } else
4992 return ERR_PTR(-ENOMEM);
4993}
4994
4995
4996static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4997{
4998 switch (algo) {
4999 case ALGORITHM_PARITY_0:
5000 if (raid_disk < max_degraded)
5001 return 1;
5002 break;
5003 case ALGORITHM_PARITY_N:
5004 if (raid_disk >= raid_disks - max_degraded)
5005 return 1;
5006 break;
5007 case ALGORITHM_PARITY_0_6:
5008 if (raid_disk == 0 ||
5009 raid_disk == raid_disks - 1)
5010 return 1;
5011 break;
5012 case ALGORITHM_LEFT_ASYMMETRIC_6:
5013 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5014 case ALGORITHM_LEFT_SYMMETRIC_6:
5015 case ALGORITHM_RIGHT_SYMMETRIC_6:
5016 if (raid_disk == raid_disks - 1)
5017 return 1;
5018 }
5019 return 0;
5020}
5021
5022static int run(struct mddev *mddev)
5023{
5024 struct r5conf *conf;
5025 int working_disks = 0;
5026 int dirty_parity_disks = 0;
5027 struct md_rdev *rdev;
5028 sector_t reshape_offset = 0;
5029 int i;
5030 long long min_offset_diff = 0;
5031 int first = 1;
5032
5033 if (mddev->recovery_cp != MaxSector)
5034 printk(KERN_NOTICE "md/raid:%s: not clean"
5035 " -- starting background reconstruction\n",
5036 mdname(mddev));
5037
5038 rdev_for_each(rdev, mddev) {
5039 long long diff;
5040 if (rdev->raid_disk < 0)
5041 continue;
5042 diff = (rdev->new_data_offset - rdev->data_offset);
5043 if (first) {
5044 min_offset_diff = diff;
5045 first = 0;
5046 } else if (mddev->reshape_backwards &&
5047 diff < min_offset_diff)
5048 min_offset_diff = diff;
5049 else if (!mddev->reshape_backwards &&
5050 diff > min_offset_diff)
5051 min_offset_diff = diff;
5052 }
5053
5054 if (mddev->reshape_position != MaxSector) {
5055 /* Check that we can continue the reshape.
5056 * Difficulties arise if the stripe we would write to
5057 * next is at or after the stripe we would read from next.
5058 * For a reshape that changes the number of devices, this
5059 * is only possible for a very short time, and mdadm makes
5060 * sure that time appears to have past before assembling
5061 * the array. So we fail if that time hasn't passed.
5062 * For a reshape that keeps the number of devices the same
5063 * mdadm must be monitoring the reshape can keeping the
5064 * critical areas read-only and backed up. It will start
5065 * the array in read-only mode, so we check for that.
5066 */
5067 sector_t here_new, here_old;
5068 int old_disks;
5069 int max_degraded = (mddev->level == 6 ? 2 : 1);
5070
5071 if (mddev->new_level != mddev->level) {
5072 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5073 "required - aborting.\n",
5074 mdname(mddev));
5075 return -EINVAL;
5076 }
5077 old_disks = mddev->raid_disks - mddev->delta_disks;
5078 /* reshape_position must be on a new-stripe boundary, and one
5079 * further up in new geometry must map after here in old
5080 * geometry.
5081 */
5082 here_new = mddev->reshape_position;
5083 if (sector_div(here_new, mddev->new_chunk_sectors *
5084 (mddev->raid_disks - max_degraded))) {
5085 printk(KERN_ERR "md/raid:%s: reshape_position not "
5086 "on a stripe boundary\n", mdname(mddev));
5087 return -EINVAL;
5088 }
5089 reshape_offset = here_new * mddev->new_chunk_sectors;
5090 /* here_new is the stripe we will write to */
5091 here_old = mddev->reshape_position;
5092 sector_div(here_old, mddev->chunk_sectors *
5093 (old_disks-max_degraded));
5094 /* here_old is the first stripe that we might need to read
5095 * from */
5096 if (mddev->delta_disks == 0) {
5097 if ((here_new * mddev->new_chunk_sectors !=
5098 here_old * mddev->chunk_sectors)) {
5099 printk(KERN_ERR "md/raid:%s: reshape position is"
5100 " confused - aborting\n", mdname(mddev));
5101 return -EINVAL;
5102 }
5103 /* We cannot be sure it is safe to start an in-place
5104 * reshape. It is only safe if user-space is monitoring
5105 * and taking constant backups.
5106 * mdadm always starts a situation like this in
5107 * readonly mode so it can take control before
5108 * allowing any writes. So just check for that.
5109 */
5110 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5111 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5112 /* not really in-place - so OK */;
5113 else if (mddev->ro == 0) {
5114 printk(KERN_ERR "md/raid:%s: in-place reshape "
5115 "must be started in read-only mode "
5116 "- aborting\n",
5117 mdname(mddev));
5118 return -EINVAL;
5119 }
5120 } else if (mddev->reshape_backwards
5121 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5122 here_old * mddev->chunk_sectors)
5123 : (here_new * mddev->new_chunk_sectors >=
5124 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5125 /* Reading from the same stripe as writing to - bad */
5126 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5127 "auto-recovery - aborting.\n",
5128 mdname(mddev));
5129 return -EINVAL;
5130 }
5131 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5132 mdname(mddev));
5133 /* OK, we should be able to continue; */
5134 } else {
5135 BUG_ON(mddev->level != mddev->new_level);
5136 BUG_ON(mddev->layout != mddev->new_layout);
5137 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5138 BUG_ON(mddev->delta_disks != 0);
5139 }
5140
5141 if (mddev->private == NULL)
5142 conf = setup_conf(mddev);
5143 else
5144 conf = mddev->private;
5145
5146 if (IS_ERR(conf))
5147 return PTR_ERR(conf);
5148
5149 conf->min_offset_diff = min_offset_diff;
5150 mddev->thread = conf->thread;
5151 conf->thread = NULL;
5152 mddev->private = conf;
5153
5154 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5155 i++) {
5156 rdev = conf->disks[i].rdev;
5157 if (!rdev && conf->disks[i].replacement) {
5158 /* The replacement is all we have yet */
5159 rdev = conf->disks[i].replacement;
5160 conf->disks[i].replacement = NULL;
5161 clear_bit(Replacement, &rdev->flags);
5162 conf->disks[i].rdev = rdev;
5163 }
5164 if (!rdev)
5165 continue;
5166 if (conf->disks[i].replacement &&
5167 conf->reshape_progress != MaxSector) {
5168 /* replacements and reshape simply do not mix. */
5169 printk(KERN_ERR "md: cannot handle concurrent "
5170 "replacement and reshape.\n");
5171 goto abort;
5172 }
5173 if (test_bit(In_sync, &rdev->flags)) {
5174 working_disks++;
5175 continue;
5176 }
5177 /* This disc is not fully in-sync. However if it
5178 * just stored parity (beyond the recovery_offset),
5179 * when we don't need to be concerned about the
5180 * array being dirty.
5181 * When reshape goes 'backwards', we never have
5182 * partially completed devices, so we only need
5183 * to worry about reshape going forwards.
5184 */
5185 /* Hack because v0.91 doesn't store recovery_offset properly. */
5186 if (mddev->major_version == 0 &&
5187 mddev->minor_version > 90)
5188 rdev->recovery_offset = reshape_offset;
5189
5190 if (rdev->recovery_offset < reshape_offset) {
5191 /* We need to check old and new layout */
5192 if (!only_parity(rdev->raid_disk,
5193 conf->algorithm,
5194 conf->raid_disks,
5195 conf->max_degraded))
5196 continue;
5197 }
5198 if (!only_parity(rdev->raid_disk,
5199 conf->prev_algo,
5200 conf->previous_raid_disks,
5201 conf->max_degraded))
5202 continue;
5203 dirty_parity_disks++;
5204 }
5205
5206 /*
5207 * 0 for a fully functional array, 1 or 2 for a degraded array.
5208 */
5209 mddev->degraded = calc_degraded(conf);
5210
5211 if (has_failed(conf)) {
5212 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5213 " (%d/%d failed)\n",
5214 mdname(mddev), mddev->degraded, conf->raid_disks);
5215 goto abort;
5216 }
5217
5218 /* device size must be a multiple of chunk size */
5219 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5220 mddev->resync_max_sectors = mddev->dev_sectors;
5221
5222 if (mddev->degraded > dirty_parity_disks &&
5223 mddev->recovery_cp != MaxSector) {
5224 if (mddev->ok_start_degraded)
5225 printk(KERN_WARNING
5226 "md/raid:%s: starting dirty degraded array"
5227 " - data corruption possible.\n",
5228 mdname(mddev));
5229 else {
5230 printk(KERN_ERR
5231 "md/raid:%s: cannot start dirty degraded array.\n",
5232 mdname(mddev));
5233 goto abort;
5234 }
5235 }
5236
5237 if (mddev->degraded == 0)
5238 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5239 " devices, algorithm %d\n", mdname(mddev), conf->level,
5240 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5241 mddev->new_layout);
5242 else
5243 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5244 " out of %d devices, algorithm %d\n",
5245 mdname(mddev), conf->level,
5246 mddev->raid_disks - mddev->degraded,
5247 mddev->raid_disks, mddev->new_layout);
5248
5249 print_raid5_conf(conf);
5250
5251 if (conf->reshape_progress != MaxSector) {
5252 conf->reshape_safe = conf->reshape_progress;
5253 atomic_set(&conf->reshape_stripes, 0);
5254 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5255 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5256 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5257 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5258 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5259 "reshape");
5260 }
5261
5262
5263 /* Ok, everything is just fine now */
5264 if (mddev->to_remove == &raid5_attrs_group)
5265 mddev->to_remove = NULL;
5266 else if (mddev->kobj.sd &&
5267 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5268 printk(KERN_WARNING
5269 "raid5: failed to create sysfs attributes for %s\n",
5270 mdname(mddev));
5271 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5272
5273 if (mddev->queue) {
5274 int chunk_size;
5275 /* read-ahead size must cover two whole stripes, which
5276 * is 2 * (datadisks) * chunksize where 'n' is the
5277 * number of raid devices
5278 */
5279 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5280 int stripe = data_disks *
5281 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5282 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5283 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5284
5285 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5286
5287 mddev->queue->backing_dev_info.congested_data = mddev;
5288 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5289
5290 chunk_size = mddev->chunk_sectors << 9;
5291 blk_queue_io_min(mddev->queue, chunk_size);
5292 blk_queue_io_opt(mddev->queue, chunk_size *
5293 (conf->raid_disks - conf->max_degraded));
5294
5295 rdev_for_each(rdev, mddev) {
5296 disk_stack_limits(mddev->gendisk, rdev->bdev,
5297 rdev->data_offset << 9);
5298 disk_stack_limits(mddev->gendisk, rdev->bdev,
5299 rdev->new_data_offset << 9);
5300 }
5301 }
5302
5303 return 0;
5304abort:
5305 md_unregister_thread(&mddev->thread);
5306 print_raid5_conf(conf);
5307 free_conf(conf);
5308 mddev->private = NULL;
5309 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5310 return -EIO;
5311}
5312
5313static int stop(struct mddev *mddev)
5314{
5315 struct r5conf *conf = mddev->private;
5316
5317 md_unregister_thread(&mddev->thread);
5318 if (mddev->queue)
5319 mddev->queue->backing_dev_info.congested_fn = NULL;
5320 free_conf(conf);
5321 mddev->private = NULL;
5322 mddev->to_remove = &raid5_attrs_group;
5323 return 0;
5324}
5325
5326static void status(struct seq_file *seq, struct mddev *mddev)
5327{
5328 struct r5conf *conf = mddev->private;
5329 int i;
5330
5331 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5332 mddev->chunk_sectors / 2, mddev->layout);
5333 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5334 for (i = 0; i < conf->raid_disks; i++)
5335 seq_printf (seq, "%s",
5336 conf->disks[i].rdev &&
5337 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5338 seq_printf (seq, "]");
5339}
5340
5341static void print_raid5_conf (struct r5conf *conf)
5342{
5343 int i;
5344 struct disk_info *tmp;
5345
5346 printk(KERN_DEBUG "RAID conf printout:\n");
5347 if (!conf) {
5348 printk("(conf==NULL)\n");
5349 return;
5350 }
5351 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5352 conf->raid_disks,
5353 conf->raid_disks - conf->mddev->degraded);
5354
5355 for (i = 0; i < conf->raid_disks; i++) {
5356 char b[BDEVNAME_SIZE];
5357 tmp = conf->disks + i;
5358 if (tmp->rdev)
5359 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5360 i, !test_bit(Faulty, &tmp->rdev->flags),
5361 bdevname(tmp->rdev->bdev, b));
5362 }
5363}
5364
5365static int raid5_spare_active(struct mddev *mddev)
5366{
5367 int i;
5368 struct r5conf *conf = mddev->private;
5369 struct disk_info *tmp;
5370 int count = 0;
5371 unsigned long flags;
5372
5373 for (i = 0; i < conf->raid_disks; i++) {
5374 tmp = conf->disks + i;
5375 if (tmp->replacement
5376 && tmp->replacement->recovery_offset == MaxSector
5377 && !test_bit(Faulty, &tmp->replacement->flags)
5378 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5379 /* Replacement has just become active. */
5380 if (!tmp->rdev
5381 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5382 count++;
5383 if (tmp->rdev) {
5384 /* Replaced device not technically faulty,
5385 * but we need to be sure it gets removed
5386 * and never re-added.
5387 */
5388 set_bit(Faulty, &tmp->rdev->flags);
5389 sysfs_notify_dirent_safe(
5390 tmp->rdev->sysfs_state);
5391 }
5392 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5393 } else if (tmp->rdev
5394 && tmp->rdev->recovery_offset == MaxSector
5395 && !test_bit(Faulty, &tmp->rdev->flags)
5396 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5397 count++;
5398 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5399 }
5400 }
5401 spin_lock_irqsave(&conf->device_lock, flags);
5402 mddev->degraded = calc_degraded(conf);
5403 spin_unlock_irqrestore(&conf->device_lock, flags);
5404 print_raid5_conf(conf);
5405 return count;
5406}
5407
5408static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5409{
5410 struct r5conf *conf = mddev->private;
5411 int err = 0;
5412 int number = rdev->raid_disk;
5413 struct md_rdev **rdevp;
5414 struct disk_info *p = conf->disks + number;
5415
5416 print_raid5_conf(conf);
5417 if (rdev == p->rdev)
5418 rdevp = &p->rdev;
5419 else if (rdev == p->replacement)
5420 rdevp = &p->replacement;
5421 else
5422 return 0;
5423
5424 if (number >= conf->raid_disks &&
5425 conf->reshape_progress == MaxSector)
5426 clear_bit(In_sync, &rdev->flags);
5427
5428 if (test_bit(In_sync, &rdev->flags) ||
5429 atomic_read(&rdev->nr_pending)) {
5430 err = -EBUSY;
5431 goto abort;
5432 }
5433 /* Only remove non-faulty devices if recovery
5434 * isn't possible.
5435 */
5436 if (!test_bit(Faulty, &rdev->flags) &&
5437 mddev->recovery_disabled != conf->recovery_disabled &&
5438 !has_failed(conf) &&
5439 (!p->replacement || p->replacement == rdev) &&
5440 number < conf->raid_disks) {
5441 err = -EBUSY;
5442 goto abort;
5443 }
5444 *rdevp = NULL;
5445 synchronize_rcu();
5446 if (atomic_read(&rdev->nr_pending)) {
5447 /* lost the race, try later */
5448 err = -EBUSY;
5449 *rdevp = rdev;
5450 } else if (p->replacement) {
5451 /* We must have just cleared 'rdev' */
5452 p->rdev = p->replacement;
5453 clear_bit(Replacement, &p->replacement->flags);
5454 smp_mb(); /* Make sure other CPUs may see both as identical
5455 * but will never see neither - if they are careful
5456 */
5457 p->replacement = NULL;
5458 clear_bit(WantReplacement, &rdev->flags);
5459 } else
5460 /* We might have just removed the Replacement as faulty-
5461 * clear the bit just in case
5462 */
5463 clear_bit(WantReplacement, &rdev->flags);
5464abort:
5465
5466 print_raid5_conf(conf);
5467 return err;
5468}
5469
5470static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5471{
5472 struct r5conf *conf = mddev->private;
5473 int err = -EEXIST;
5474 int disk;
5475 struct disk_info *p;
5476 int first = 0;
5477 int last = conf->raid_disks - 1;
5478
5479 if (mddev->recovery_disabled == conf->recovery_disabled)
5480 return -EBUSY;
5481
5482 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5483 /* no point adding a device */
5484 return -EINVAL;
5485
5486 if (rdev->raid_disk >= 0)
5487 first = last = rdev->raid_disk;
5488
5489 /*
5490 * find the disk ... but prefer rdev->saved_raid_disk
5491 * if possible.
5492 */
5493 if (rdev->saved_raid_disk >= 0 &&
5494 rdev->saved_raid_disk >= first &&
5495 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5496 first = rdev->saved_raid_disk;
5497
5498 for (disk = first; disk <= last; disk++) {
5499 p = conf->disks + disk;
5500 if (p->rdev == NULL) {
5501 clear_bit(In_sync, &rdev->flags);
5502 rdev->raid_disk = disk;
5503 err = 0;
5504 if (rdev->saved_raid_disk != disk)
5505 conf->fullsync = 1;
5506 rcu_assign_pointer(p->rdev, rdev);
5507 goto out;
5508 }
5509 }
5510 for (disk = first; disk <= last; disk++) {
5511 p = conf->disks + disk;
5512 if (test_bit(WantReplacement, &p->rdev->flags) &&
5513 p->replacement == NULL) {
5514 clear_bit(In_sync, &rdev->flags);
5515 set_bit(Replacement, &rdev->flags);
5516 rdev->raid_disk = disk;
5517 err = 0;
5518 conf->fullsync = 1;
5519 rcu_assign_pointer(p->replacement, rdev);
5520 break;
5521 }
5522 }
5523out:
5524 print_raid5_conf(conf);
5525 return err;
5526}
5527
5528static int raid5_resize(struct mddev *mddev, sector_t sectors)
5529{
5530 /* no resync is happening, and there is enough space
5531 * on all devices, so we can resize.
5532 * We need to make sure resync covers any new space.
5533 * If the array is shrinking we should possibly wait until
5534 * any io in the removed space completes, but it hardly seems
5535 * worth it.
5536 */
5537 sector_t newsize;
5538 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5539 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5540 if (mddev->external_size &&
5541 mddev->array_sectors > newsize)
5542 return -EINVAL;
5543 if (mddev->bitmap) {
5544 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5545 if (ret)
5546 return ret;
5547 }
5548 md_set_array_sectors(mddev, newsize);
5549 set_capacity(mddev->gendisk, mddev->array_sectors);
5550 revalidate_disk(mddev->gendisk);
5551 if (sectors > mddev->dev_sectors &&
5552 mddev->recovery_cp > mddev->dev_sectors) {
5553 mddev->recovery_cp = mddev->dev_sectors;
5554 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5555 }
5556 mddev->dev_sectors = sectors;
5557 mddev->resync_max_sectors = sectors;
5558 return 0;
5559}
5560
5561static int check_stripe_cache(struct mddev *mddev)
5562{
5563 /* Can only proceed if there are plenty of stripe_heads.
5564 * We need a minimum of one full stripe,, and for sensible progress
5565 * it is best to have about 4 times that.
5566 * If we require 4 times, then the default 256 4K stripe_heads will
5567 * allow for chunk sizes up to 256K, which is probably OK.
5568 * If the chunk size is greater, user-space should request more
5569 * stripe_heads first.
5570 */
5571 struct r5conf *conf = mddev->private;
5572 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5573 > conf->max_nr_stripes ||
5574 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5575 > conf->max_nr_stripes) {
5576 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5577 mdname(mddev),
5578 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5579 / STRIPE_SIZE)*4);
5580 return 0;
5581 }
5582 return 1;
5583}
5584
5585static int check_reshape(struct mddev *mddev)
5586{
5587 struct r5conf *conf = mddev->private;
5588
5589 if (mddev->delta_disks == 0 &&
5590 mddev->new_layout == mddev->layout &&
5591 mddev->new_chunk_sectors == mddev->chunk_sectors)
5592 return 0; /* nothing to do */
5593 if (has_failed(conf))
5594 return -EINVAL;
5595 if (mddev->delta_disks < 0) {
5596 /* We might be able to shrink, but the devices must
5597 * be made bigger first.
5598 * For raid6, 4 is the minimum size.
5599 * Otherwise 2 is the minimum
5600 */
5601 int min = 2;
5602 if (mddev->level == 6)
5603 min = 4;
5604 if (mddev->raid_disks + mddev->delta_disks < min)
5605 return -EINVAL;
5606 }
5607
5608 if (!check_stripe_cache(mddev))
5609 return -ENOSPC;
5610
5611 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5612}
5613
5614static int raid5_start_reshape(struct mddev *mddev)
5615{
5616 struct r5conf *conf = mddev->private;
5617 struct md_rdev *rdev;
5618 int spares = 0;
5619 unsigned long flags;
5620
5621 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5622 return -EBUSY;
5623
5624 if (!check_stripe_cache(mddev))
5625 return -ENOSPC;
5626
5627 if (has_failed(conf))
5628 return -EINVAL;
5629
5630 rdev_for_each(rdev, mddev) {
5631 if (!test_bit(In_sync, &rdev->flags)
5632 && !test_bit(Faulty, &rdev->flags))
5633 spares++;
5634 }
5635
5636 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5637 /* Not enough devices even to make a degraded array
5638 * of that size
5639 */
5640 return -EINVAL;
5641
5642 /* Refuse to reduce size of the array. Any reductions in
5643 * array size must be through explicit setting of array_size
5644 * attribute.
5645 */
5646 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5647 < mddev->array_sectors) {
5648 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5649 "before number of disks\n", mdname(mddev));
5650 return -EINVAL;
5651 }
5652
5653 atomic_set(&conf->reshape_stripes, 0);
5654 spin_lock_irq(&conf->device_lock);
5655 conf->previous_raid_disks = conf->raid_disks;
5656 conf->raid_disks += mddev->delta_disks;
5657 conf->prev_chunk_sectors = conf->chunk_sectors;
5658 conf->chunk_sectors = mddev->new_chunk_sectors;
5659 conf->prev_algo = conf->algorithm;
5660 conf->algorithm = mddev->new_layout;
5661 conf->generation++;
5662 /* Code that selects data_offset needs to see the generation update
5663 * if reshape_progress has been set - so a memory barrier needed.
5664 */
5665 smp_mb();
5666 if (mddev->reshape_backwards)
5667 conf->reshape_progress = raid5_size(mddev, 0, 0);
5668 else
5669 conf->reshape_progress = 0;
5670 conf->reshape_safe = conf->reshape_progress;
5671 spin_unlock_irq(&conf->device_lock);
5672
5673 /* Add some new drives, as many as will fit.
5674 * We know there are enough to make the newly sized array work.
5675 * Don't add devices if we are reducing the number of
5676 * devices in the array. This is because it is not possible
5677 * to correctly record the "partially reconstructed" state of
5678 * such devices during the reshape and confusion could result.
5679 */
5680 if (mddev->delta_disks >= 0) {
5681 rdev_for_each(rdev, mddev)
5682 if (rdev->raid_disk < 0 &&
5683 !test_bit(Faulty, &rdev->flags)) {
5684 if (raid5_add_disk(mddev, rdev) == 0) {
5685 if (rdev->raid_disk
5686 >= conf->previous_raid_disks)
5687 set_bit(In_sync, &rdev->flags);
5688 else
5689 rdev->recovery_offset = 0;
5690
5691 if (sysfs_link_rdev(mddev, rdev))
5692 /* Failure here is OK */;
5693 }
5694 } else if (rdev->raid_disk >= conf->previous_raid_disks
5695 && !test_bit(Faulty, &rdev->flags)) {
5696 /* This is a spare that was manually added */
5697 set_bit(In_sync, &rdev->flags);
5698 }
5699
5700 /* When a reshape changes the number of devices,
5701 * ->degraded is measured against the larger of the
5702 * pre and post number of devices.
5703 */
5704 spin_lock_irqsave(&conf->device_lock, flags);
5705 mddev->degraded = calc_degraded(conf);
5706 spin_unlock_irqrestore(&conf->device_lock, flags);
5707 }
5708 mddev->raid_disks = conf->raid_disks;
5709 mddev->reshape_position = conf->reshape_progress;
5710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5711
5712 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5713 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5714 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5715 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5716 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5717 "reshape");
5718 if (!mddev->sync_thread) {
5719 mddev->recovery = 0;
5720 spin_lock_irq(&conf->device_lock);
5721 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5722 rdev_for_each(rdev, mddev)
5723 rdev->new_data_offset = rdev->data_offset;
5724 smp_wmb();
5725 conf->reshape_progress = MaxSector;
5726 mddev->reshape_position = MaxSector;
5727 spin_unlock_irq(&conf->device_lock);
5728 return -EAGAIN;
5729 }
5730 conf->reshape_checkpoint = jiffies;
5731 md_wakeup_thread(mddev->sync_thread);
5732 md_new_event(mddev);
5733 return 0;
5734}
5735
5736/* This is called from the reshape thread and should make any
5737 * changes needed in 'conf'
5738 */
5739static void end_reshape(struct r5conf *conf)
5740{
5741
5742 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5743 struct md_rdev *rdev;
5744
5745 spin_lock_irq(&conf->device_lock);
5746 conf->previous_raid_disks = conf->raid_disks;
5747 rdev_for_each(rdev, conf->mddev)
5748 rdev->data_offset = rdev->new_data_offset;
5749 smp_wmb();
5750 conf->reshape_progress = MaxSector;
5751 spin_unlock_irq(&conf->device_lock);
5752 wake_up(&conf->wait_for_overlap);
5753
5754 /* read-ahead size must cover two whole stripes, which is
5755 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5756 */
5757 if (conf->mddev->queue) {
5758 int data_disks = conf->raid_disks - conf->max_degraded;
5759 int stripe = data_disks * ((conf->chunk_sectors << 9)
5760 / PAGE_SIZE);
5761 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5762 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5763 }
5764 }
5765}
5766
5767/* This is called from the raid5d thread with mddev_lock held.
5768 * It makes config changes to the device.
5769 */
5770static void raid5_finish_reshape(struct mddev *mddev)
5771{
5772 struct r5conf *conf = mddev->private;
5773
5774 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5775
5776 if (mddev->delta_disks > 0) {
5777 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5778 set_capacity(mddev->gendisk, mddev->array_sectors);
5779 revalidate_disk(mddev->gendisk);
5780 } else {
5781 int d;
5782 spin_lock_irq(&conf->device_lock);
5783 mddev->degraded = calc_degraded(conf);
5784 spin_unlock_irq(&conf->device_lock);
5785 for (d = conf->raid_disks ;
5786 d < conf->raid_disks - mddev->delta_disks;
5787 d++) {
5788 struct md_rdev *rdev = conf->disks[d].rdev;
5789 if (rdev)
5790 clear_bit(In_sync, &rdev->flags);
5791 rdev = conf->disks[d].replacement;
5792 if (rdev)
5793 clear_bit(In_sync, &rdev->flags);
5794 }
5795 }
5796 mddev->layout = conf->algorithm;
5797 mddev->chunk_sectors = conf->chunk_sectors;
5798 mddev->reshape_position = MaxSector;
5799 mddev->delta_disks = 0;
5800 mddev->reshape_backwards = 0;
5801 }
5802}
5803
5804static void raid5_quiesce(struct mddev *mddev, int state)
5805{
5806 struct r5conf *conf = mddev->private;
5807
5808 switch(state) {
5809 case 2: /* resume for a suspend */
5810 wake_up(&conf->wait_for_overlap);
5811 break;
5812
5813 case 1: /* stop all writes */
5814 spin_lock_irq(&conf->device_lock);
5815 /* '2' tells resync/reshape to pause so that all
5816 * active stripes can drain
5817 */
5818 conf->quiesce = 2;
5819 wait_event_lock_irq(conf->wait_for_stripe,
5820 atomic_read(&conf->active_stripes) == 0 &&
5821 atomic_read(&conf->active_aligned_reads) == 0,
5822 conf->device_lock, /* nothing */);
5823 conf->quiesce = 1;
5824 spin_unlock_irq(&conf->device_lock);
5825 /* allow reshape to continue */
5826 wake_up(&conf->wait_for_overlap);
5827 break;
5828
5829 case 0: /* re-enable writes */
5830 spin_lock_irq(&conf->device_lock);
5831 conf->quiesce = 0;
5832 wake_up(&conf->wait_for_stripe);
5833 wake_up(&conf->wait_for_overlap);
5834 spin_unlock_irq(&conf->device_lock);
5835 break;
5836 }
5837}
5838
5839
5840static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5841{
5842 struct r0conf *raid0_conf = mddev->private;
5843 sector_t sectors;
5844
5845 /* for raid0 takeover only one zone is supported */
5846 if (raid0_conf->nr_strip_zones > 1) {
5847 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5848 mdname(mddev));
5849 return ERR_PTR(-EINVAL);
5850 }
5851
5852 sectors = raid0_conf->strip_zone[0].zone_end;
5853 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5854 mddev->dev_sectors = sectors;
5855 mddev->new_level = level;
5856 mddev->new_layout = ALGORITHM_PARITY_N;
5857 mddev->new_chunk_sectors = mddev->chunk_sectors;
5858 mddev->raid_disks += 1;
5859 mddev->delta_disks = 1;
5860 /* make sure it will be not marked as dirty */
5861 mddev->recovery_cp = MaxSector;
5862
5863 return setup_conf(mddev);
5864}
5865
5866
5867static void *raid5_takeover_raid1(struct mddev *mddev)
5868{
5869 int chunksect;
5870
5871 if (mddev->raid_disks != 2 ||
5872 mddev->degraded > 1)
5873 return ERR_PTR(-EINVAL);
5874
5875 /* Should check if there are write-behind devices? */
5876
5877 chunksect = 64*2; /* 64K by default */
5878
5879 /* The array must be an exact multiple of chunksize */
5880 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5881 chunksect >>= 1;
5882
5883 if ((chunksect<<9) < STRIPE_SIZE)
5884 /* array size does not allow a suitable chunk size */
5885 return ERR_PTR(-EINVAL);
5886
5887 mddev->new_level = 5;
5888 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5889 mddev->new_chunk_sectors = chunksect;
5890
5891 return setup_conf(mddev);
5892}
5893
5894static void *raid5_takeover_raid6(struct mddev *mddev)
5895{
5896 int new_layout;
5897
5898 switch (mddev->layout) {
5899 case ALGORITHM_LEFT_ASYMMETRIC_6:
5900 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5901 break;
5902 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5903 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5904 break;
5905 case ALGORITHM_LEFT_SYMMETRIC_6:
5906 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5907 break;
5908 case ALGORITHM_RIGHT_SYMMETRIC_6:
5909 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5910 break;
5911 case ALGORITHM_PARITY_0_6:
5912 new_layout = ALGORITHM_PARITY_0;
5913 break;
5914 case ALGORITHM_PARITY_N:
5915 new_layout = ALGORITHM_PARITY_N;
5916 break;
5917 default:
5918 return ERR_PTR(-EINVAL);
5919 }
5920 mddev->new_level = 5;
5921 mddev->new_layout = new_layout;
5922 mddev->delta_disks = -1;
5923 mddev->raid_disks -= 1;
5924 return setup_conf(mddev);
5925}
5926
5927
5928static int raid5_check_reshape(struct mddev *mddev)
5929{
5930 /* For a 2-drive array, the layout and chunk size can be changed
5931 * immediately as not restriping is needed.
5932 * For larger arrays we record the new value - after validation
5933 * to be used by a reshape pass.
5934 */
5935 struct r5conf *conf = mddev->private;
5936 int new_chunk = mddev->new_chunk_sectors;
5937
5938 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5939 return -EINVAL;
5940 if (new_chunk > 0) {
5941 if (!is_power_of_2(new_chunk))
5942 return -EINVAL;
5943 if (new_chunk < (PAGE_SIZE>>9))
5944 return -EINVAL;
5945 if (mddev->array_sectors & (new_chunk-1))
5946 /* not factor of array size */
5947 return -EINVAL;
5948 }
5949
5950 /* They look valid */
5951
5952 if (mddev->raid_disks == 2) {
5953 /* can make the change immediately */
5954 if (mddev->new_layout >= 0) {
5955 conf->algorithm = mddev->new_layout;
5956 mddev->layout = mddev->new_layout;
5957 }
5958 if (new_chunk > 0) {
5959 conf->chunk_sectors = new_chunk ;
5960 mddev->chunk_sectors = new_chunk;
5961 }
5962 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5963 md_wakeup_thread(mddev->thread);
5964 }
5965 return check_reshape(mddev);
5966}
5967
5968static int raid6_check_reshape(struct mddev *mddev)
5969{
5970 int new_chunk = mddev->new_chunk_sectors;
5971
5972 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5973 return -EINVAL;
5974 if (new_chunk > 0) {
5975 if (!is_power_of_2(new_chunk))
5976 return -EINVAL;
5977 if (new_chunk < (PAGE_SIZE >> 9))
5978 return -EINVAL;
5979 if (mddev->array_sectors & (new_chunk-1))
5980 /* not factor of array size */
5981 return -EINVAL;
5982 }
5983
5984 /* They look valid */
5985 return check_reshape(mddev);
5986}
5987
5988static void *raid5_takeover(struct mddev *mddev)
5989{
5990 /* raid5 can take over:
5991 * raid0 - if there is only one strip zone - make it a raid4 layout
5992 * raid1 - if there are two drives. We need to know the chunk size
5993 * raid4 - trivial - just use a raid4 layout.
5994 * raid6 - Providing it is a *_6 layout
5995 */
5996 if (mddev->level == 0)
5997 return raid45_takeover_raid0(mddev, 5);
5998 if (mddev->level == 1)
5999 return raid5_takeover_raid1(mddev);
6000 if (mddev->level == 4) {
6001 mddev->new_layout = ALGORITHM_PARITY_N;
6002 mddev->new_level = 5;
6003 return setup_conf(mddev);
6004 }
6005 if (mddev->level == 6)
6006 return raid5_takeover_raid6(mddev);
6007
6008 return ERR_PTR(-EINVAL);
6009}
6010
6011static void *raid4_takeover(struct mddev *mddev)
6012{
6013 /* raid4 can take over:
6014 * raid0 - if there is only one strip zone
6015 * raid5 - if layout is right
6016 */
6017 if (mddev->level == 0)
6018 return raid45_takeover_raid0(mddev, 4);
6019 if (mddev->level == 5 &&
6020 mddev->layout == ALGORITHM_PARITY_N) {
6021 mddev->new_layout = 0;
6022 mddev->new_level = 4;
6023 return setup_conf(mddev);
6024 }
6025 return ERR_PTR(-EINVAL);
6026}
6027
6028static struct md_personality raid5_personality;
6029
6030static void *raid6_takeover(struct mddev *mddev)
6031{
6032 /* Currently can only take over a raid5. We map the
6033 * personality to an equivalent raid6 personality
6034 * with the Q block at the end.
6035 */
6036 int new_layout;
6037
6038 if (mddev->pers != &raid5_personality)
6039 return ERR_PTR(-EINVAL);
6040 if (mddev->degraded > 1)
6041 return ERR_PTR(-EINVAL);
6042 if (mddev->raid_disks > 253)
6043 return ERR_PTR(-EINVAL);
6044 if (mddev->raid_disks < 3)
6045 return ERR_PTR(-EINVAL);
6046
6047 switch (mddev->layout) {
6048 case ALGORITHM_LEFT_ASYMMETRIC:
6049 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6050 break;
6051 case ALGORITHM_RIGHT_ASYMMETRIC:
6052 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6053 break;
6054 case ALGORITHM_LEFT_SYMMETRIC:
6055 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6056 break;
6057 case ALGORITHM_RIGHT_SYMMETRIC:
6058 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6059 break;
6060 case ALGORITHM_PARITY_0:
6061 new_layout = ALGORITHM_PARITY_0_6;
6062 break;
6063 case ALGORITHM_PARITY_N:
6064 new_layout = ALGORITHM_PARITY_N;
6065 break;
6066 default:
6067 return ERR_PTR(-EINVAL);
6068 }
6069 mddev->new_level = 6;
6070 mddev->new_layout = new_layout;
6071 mddev->delta_disks = 1;
6072 mddev->raid_disks += 1;
6073 return setup_conf(mddev);
6074}
6075
6076
6077static struct md_personality raid6_personality =
6078{
6079 .name = "raid6",
6080 .level = 6,
6081 .owner = THIS_MODULE,
6082 .make_request = make_request,
6083 .run = run,
6084 .stop = stop,
6085 .status = status,
6086 .error_handler = error,
6087 .hot_add_disk = raid5_add_disk,
6088 .hot_remove_disk= raid5_remove_disk,
6089 .spare_active = raid5_spare_active,
6090 .sync_request = sync_request,
6091 .resize = raid5_resize,
6092 .size = raid5_size,
6093 .check_reshape = raid6_check_reshape,
6094 .start_reshape = raid5_start_reshape,
6095 .finish_reshape = raid5_finish_reshape,
6096 .quiesce = raid5_quiesce,
6097 .takeover = raid6_takeover,
6098};
6099static struct md_personality raid5_personality =
6100{
6101 .name = "raid5",
6102 .level = 5,
6103 .owner = THIS_MODULE,
6104 .make_request = make_request,
6105 .run = run,
6106 .stop = stop,
6107 .status = status,
6108 .error_handler = error,
6109 .hot_add_disk = raid5_add_disk,
6110 .hot_remove_disk= raid5_remove_disk,
6111 .spare_active = raid5_spare_active,
6112 .sync_request = sync_request,
6113 .resize = raid5_resize,
6114 .size = raid5_size,
6115 .check_reshape = raid5_check_reshape,
6116 .start_reshape = raid5_start_reshape,
6117 .finish_reshape = raid5_finish_reshape,
6118 .quiesce = raid5_quiesce,
6119 .takeover = raid5_takeover,
6120};
6121
6122static struct md_personality raid4_personality =
6123{
6124 .name = "raid4",
6125 .level = 4,
6126 .owner = THIS_MODULE,
6127 .make_request = make_request,
6128 .run = run,
6129 .stop = stop,
6130 .status = status,
6131 .error_handler = error,
6132 .hot_add_disk = raid5_add_disk,
6133 .hot_remove_disk= raid5_remove_disk,
6134 .spare_active = raid5_spare_active,
6135 .sync_request = sync_request,
6136 .resize = raid5_resize,
6137 .size = raid5_size,
6138 .check_reshape = raid5_check_reshape,
6139 .start_reshape = raid5_start_reshape,
6140 .finish_reshape = raid5_finish_reshape,
6141 .quiesce = raid5_quiesce,
6142 .takeover = raid4_takeover,
6143};
6144
6145static int __init raid5_init(void)
6146{
6147 register_md_personality(&raid6_personality);
6148 register_md_personality(&raid5_personality);
6149 register_md_personality(&raid4_personality);
6150 return 0;
6151}
6152
6153static void raid5_exit(void)
6154{
6155 unregister_md_personality(&raid6_personality);
6156 unregister_md_personality(&raid5_personality);
6157 unregister_md_personality(&raid4_personality);
6158}
6159
6160module_init(raid5_init);
6161module_exit(raid5_exit);
6162MODULE_LICENSE("GPL");
6163MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6164MODULE_ALIAS("md-personality-4"); /* RAID5 */
6165MODULE_ALIAS("md-raid5");
6166MODULE_ALIAS("md-raid4");
6167MODULE_ALIAS("md-level-5");
6168MODULE_ALIAS("md-level-4");
6169MODULE_ALIAS("md-personality-8"); /* RAID6 */
6170MODULE_ALIAS("md-raid6");
6171MODULE_ALIAS("md-level-6");
6172
6173/* This used to be two separate modules, they were: */
6174MODULE_ALIAS("raid5");
6175MODULE_ALIAS("raid6");